feat: generic support for "pseudo" or "hidden" columns

[jcrist]

Some databases have "hidden" columns that can be referenced in a query, but won't show up (by default) in a SELECT *. Oracle calls these "pseudocolumns", and that name seems to have stuck with other databases too. Personally I think "hidden" is a more descriptive name, but 🤷

A few examples:

• Snowflake's METADATA columns
• Bigquery's _TABLE_SUFFIX and _PARTITIONTIME (docs)
• Oracle's stuff (docs)
• SQLite's hidden columns (docs)

In #9375 a (pragmatic) hack was added to the bigquery backend to support filtering on _TABLE_SUFFIX for partitioned tables. This hack is unfortunate in a few ways:

• The schema of the table expression won't always match the resulting schema (in the common case, the ibis schema has _TABLE_SUFFIX, while the result doesn't).
• If a user intentionally tries to select _TABLE_SUFFIX, the resulting table still won't include it since we unconditionally drop the value. For example, t.select("_TABLE_SUFFIX", "a").execute() will just have "a".
• It messes up the result handling pipeline, requiring a bunch of special casing (see fix(bigquery): fix column name mismatches and support _TABLE_SUFFIX in all to_* methods #10048).

I propose we drop this special case in favor of a generic mechanism. This won't be as convenient for users trying to access _TABLE_SUFFIX, but it will be generic and less of a pain to maintain.

I think the easiest way to do this would be to add a method to Table (I'll call it hidden here, but not attached, could also be pseudo/pseudo_col/pseudocol/).

# When called, this method takes a name and an optional type.
# If no type is given, the type is `unknown` and will require a cast
# to do much with it.
t.filter(t.hidden("_TABLE_SUFFIX", "string") > "abc")

# With no type specified, defaults to unknown
t.filter(t.hidden("_TABLE_SUFFIX").cast("string") > "abc")

One nice thing about this (besides dropping the special casing) is it still allows users to include these columns in the result set if they're explicitly asked for:

expr = t.mutate(t.hidden("_TABLE_SUFFIX"))
expr.to_pandas()  # this will include _TABLE_SUFFIX, while currently we don't