Change readdir from returning bytes to returning a dir-entrys.

`readdir` returning bytes has been a source of complexity for both libc
and engine implementors, with subtle issues about alignment and partial
records.

Rename `dirent` to `dir-entry` for readability, and revamp it:
 - Make the name field a proper `string` rather than being represented
   by trailing bytes in the buffer.
 - Make the inode field `optional`, so that we can make it optional
   as discussed in #65 without special-casing zero, which is reportedly
   a valid inode number on [some filesystems].

And remove the `rewind` parameter, which isn't needed when we return
a stream, as users wanting to start at the beginning can just request
a new stream.

[some filesystems]: https://sourceware.org/pipermail/libc-alpha/2022-September/142059.html

wasi-filesystem.abi.md

@@ -268,25 +268,30 @@ Size: 16, Alignment: 8
 
   Set the timestamp to the given value.
 
-## <a href="#dirent" name="dirent"></a> `dirent`: record
+## <a href="#dir_entry" name="dir_entry"></a> `dir-entry`: record
 
   A directory entry.
 
-Size: 16, Alignment: 8
+Size: 32, Alignment: 8
 
 ### Record Fields
 
-- <a href="dirent.ino" name="dirent.ino"></a> [`ino`](#dirent.ino): [`inode`](#inode)
+- <a href="dir_entry.ino" name="dir_entry.ino"></a> [`ino`](#dir_entry.ino): option<[`inode`](#inode)>
 
-  The serial number of the file referred to by this directory entry.
+  The serial number of the object referred to by this directory entry.
+  May be none if the inode value is not known.
+
+  When this is none, libc implementations might do an extra `stat-at`
+  call to retrieve the inode number to fill their `d_ino` fields, so
+  implementations which can set this to a non-none value should do so.
 
-- <a href="dirent.namelen" name="dirent.namelen"></a> [`namelen`](#dirent.namelen): [`size`](#size)
+- <a href="dir_entry.type" name="dir_entry.type"></a> [`type`](#dir_entry.type): [`descriptor-type`](#descriptor_type)
 
-  The length of the name of the directory entry.
+  The type of the file referred to by this directory entry.
 
-- <a href="dirent.type" name="dirent.type"></a> [`type`](#dirent.type): [`descriptor-type`](#descriptor_type)
+- <a href="dir_entry.name" name="dir_entry.name"></a> [`name`](#dir_entry.name): `string`
 
-  The type of the file referred to by this directory entry.
+  The name of the object.
 
 ## <a href="#errno" name="errno"></a> `errno`: enum
 
@@ -787,22 +792,14 @@ Size: 16, Alignment: 8
 
   Read directory entries from a directory.
 
-  When successful, the contents of the output buffer consist of a sequence of
-  directory entries. Each directory entry consists of a `dirent` object,
-  followed by `dirent::d_namlen` bytes holding the name of the directory
-  entry.
-
-  This function fills the output buffer as much as possible, potentially
-  truncating the last directory entry. This allows the caller to grow its
-  read buffer size in case it's too small to fit a single large directory
-  entry, or skip the oversized directory entry.
+  This always returns a new stream which starts at the beginning of the
+  directory.
 ##### Params
 
 - <a href="#descriptor_readdir.self" name="descriptor_readdir.self"></a> `self`: handle<descriptor>
-- <a href="#descriptor_readdir.rewind" name="descriptor_readdir.rewind"></a> `rewind`: `bool`
 ##### Results
 
-- stream<`u8`, [`errno`](#errno)>
+- stream<[`dir-entry`](#dir_entry), [`errno`](#errno)>
 
 ----
 

wasi-filesystem.wit.md

@@ -192,16 +192,23 @@ variant new-timestamp {
 }
 ```
 
-## `dirent`
+## `dir-entry`
 ```wit
 /// A directory entry. 
-record dirent {
-    /// The serial number of the file referred to by this directory entry.
-    ino: inode,
-    /// The length of the name of the directory entry.
-    namelen: size,
+record dir-entry {
+    /// The serial number of the object referred to by this directory entry.
+    /// May be none if the inode value is not known.
+    ///
+    /// When this is none, libc implementations might do an extra `stat-at`
+    /// call to retrieve the inode number to fill their `d_ino` fields, so
+    /// implementations which can set this to a non-none value should do so.
+    ino: option<inode>,
+
     /// The type of the file referred to by this directory entry.
     %type: descriptor-type,
+
+    /// The name of the object.
+    name: string,
 }
 ```
 
@@ -506,19 +513,9 @@ pwrite: func(
 ```wit
 /// Read directory entries from a directory.
 ///
-/// When successful, the contents of the output buffer consist of a sequence of
-/// directory entries. Each directory entry consists of a `dirent` object,
-/// followed by `dirent::d_namlen` bytes holding the name of the directory
-/// entry.
-///
-/// This function fills the output buffer as much as possible, potentially
-/// truncating the last directory entry. This allows the caller to grow its
-/// read buffer size in case it's too small to fit a single large directory
-/// entry, or skip the oversized directory entry.
-readdir: func(
-    /// If true, rewind the current position to the beginning before reading.
-    rewind: bool
-) -> stream<u8, errno>
+/// This always returns a new stream which starts at the beginning of the
+/// directory.
+readdir: func() -> stream<dir-entry, errno>
 ```
 
 ## `seek`