gh-128213: fast path for bytes creation from list and tuple

[eendebakpt]

Continuation of #128214. This PR

• Improves performance of bytes creation from a list or tuple. On the FT build performance on a list of tuple of size 1000 the method is about 2 to 3x times faster (mainly because the old code uses per element atomic operations via _PyList_GetItemRef).
• Fixes a free-threading bug
• Reduces some duplicated code between list and tuple handling

• Issue: Speed up bytes creation from list and tuple #128213

[markshannon]

If you're interested in speed, PyLong_CheckExact will allow you to use _PyLong_IsNonNegativeCompact and _PyLong_CompactValue to get the C int in a few cycles.

[eendebakpt]

Interesting. For the common/happy path we now we call PyNumber_AsSsize_t, which calls PyLong_AsSsize_t. That methods does another PyLong_Check and then a call to _PyLong_IsCompact.

So we have some options:

1. We could remove the call to PyLong_Check (as it is already covered by PyLong_AsSsize_t), that improves performance a bit for all cases.
2. We could use PyLong_CheckExact with _PyLong_IsNonNegativeCompact. Fastest for exact ints, but non-exact ints become slower. I suspect non-exact ints are rare though.
3. We add a fast path using PyLong_CheckExact and a fallback fast path using PyNumber_AsSsize_t for the non-exacts ints. Fast for all cases, but takes a but more code.

@markshannon Any preference? I am happy to work out any of the above.

[eendebakpt]

The PyNumber_AsSsize_t performs an incref/decref on the argument which we can avoid by calling PyLong_AsSsize_t directly (the incref is not very bad, since in the happy path the argument will be in the 0 to 255 range so a python small int, but still). I updated the PR with this approach.

Most important for the PR is making the code thread-safe, so I left the possible optimization with _PyLong_IsNonNegativeCompact out for now.

[markshannon]

Tuples are immutable, so why does creating a bytes object from a tuple require synchronization?

[eendebakpt]

Tuples are immutable, so why does creating a bytes object from a tuple require synchronization?

Tuples indeed do not require synchronization.

In this PR exact lists and tuples use the path (using synchronization with Py_BEGIN_CRITICAL_SECTION_SEQUENCE_FAST). We could split the path, but I suspect the performance gain for tuples would be minimal.

[eendebakpt]

@markshannon Would you be able to continue reviewing this one?

[eendebakpt]

@colesbury Would you be able to review?