avoid dynamic memory allocation during error handling

[wjwwood]

This pull request refactors the way the error handling API in rcutils works so that it does not use dynamic memory and there does not need an allocator in most cases and also should make it easier to use in real-time and high performance situations.

Overview of major changes:

• rcutils_error_state_t now uses fixed length strings
• the formatted error string (rcutils_get_error_string()) returns a fixed sized string
• error string formatting no longer uses snprintf, and instead uses custom functions which explicitly avoid memory allocations
• there's no longer an rcutils_get_error_string_safe() as you cannot get a nullptr for the error string anymore (so it's always "safe")
• suppression of the warning about overriding the current error now does a string compare
  • this is ok from a performance perspective since it is only checked when the error state is set (it was not cleared before setting it again)
  • this is slightly less optimal than a direct pointer comparison because it could incorrectly suppress the warning when setting the same error message in a loop (same string, not same pointer)
• a new method rcutils_initialize_error_handling_thread_local_storage is required to be called in new threads in order to avoid memory allocations in other functions
  • it is the only function that takes an allocator now, and it is only used in certain cases
• the pthread implementation was removed
  • it complicated the implementation significantly
  • I'm not sure it needed anymore, see: Add support for pthread-based thread local storage rmw#80 (comment)
  • I'd prefer an entirely separate implementation in a separate file if we need to resurrect pthread support
• I've documented the use of dynamic memory allocation for each function and tested it with the memory testing tools in osrf_testing_tools_cpp.

Main migration issues:

• code which previously did rcutils_get_error_string_safe() will need to change to rcutils_get_error_string().str (so long as the temporary is ok)
  • to avoid the temporary, they'll need to do something like rcutils_error_string_t copy = rcutils_get_error_string();)
• the error set macros which previously required that you pass an allocator no longer takes an allocator, e.g. change RCUTILS_SET_ERROR_MSG("char array has no valid allocator", error_msg_allocator); -> RCUTILS_SET_ERROR_MSG("char array has no valid allocator");

These changes are annoying, but I think ultimately they're for the better. I considered keeping existing signatures and doing tick-tock, and that's an option of the macro's which take allocators, but not so for the get error string signature, since it needs to return the copy to the fixed sized string. I decided against doing it for the macros because it would require using a separate name for the new macros and because deprecating them is not easy in pure C.

I have a lot more changes up and down the stack due to these API changes, and I'll open pull requests for those once I get feedback on the proposed changes here.

---

/cc @serge-nikulin fyi

[wjwwood]

Here's the first "downstream" pull request due to these API changes: ros2/rmw#153

[wjwwood]

CI to see what's missing/broken:

• Linux
• Linux-aarch64
• macOS
• Windows

[clalancette]

Overall, I like the change. This initial review has some details that I came across, plus a couple of questions. I can do a bit more reviewing for it later.

[wjwwood]

New CI:

• Linux
• Linux-aarch64
• macOS
• Windows

Also, OpenSplice:

• Linux

[wjwwood]

I had to rebase a few repositories due to recent merges, so here's new CI:

• Linux
• Linux-aarch64
• macOS
• Windows

[wjwwood]

The only test failure was due to the launch problem. Also the job is marked as failing, but the job finished (no compiler error). @mjcarroll

[wjwwood]

This is waiting on ros2/launch#153 to have clean CI, but otherwise I believe CI is good for this set of pull requests and ready for review.

[jacobperron]

Some minor comments, but LGTM pending CI.

[wjwwood]

CI:

• Linux
• Linux-aarch64
• macOS
• Windows

[sloretz]

Minor comments, feel free to ignore

[wjwwood]

Ok I think I addressed all of the feedback, can you guys give it a another look.

[wjwwood]

Optimistic CI, pending feedback on the new changes:

• Linux

• Linux-aarch64

• macOS

• Windows

• Linux (OpenSplice)

[sloretz]

LGTM

[wjwwood]

Ok, I'm going to take my window to merge. I'm happy to follow up with changes in new pr's if needed.