gh-125985: Add free threading scaling micro benchmarks

[colesbury]

These consist of a number of short snippets that help identify scaling bottlenecks in the free threaded interpreter.

The current bottlenecks are in calling functions in benchmarks that call functions (due to LOAD_ATTR not yet using deferred reference counting) and when accessing thread-local data.

• Issue: Add free threading scaling microbenchmarks #125985

[colesbury]

Some results below:

CPython 3.14t results

object_cfunction 1.1x faster cmodule_function 1.1x slower mult_constant 9.7x faster generator 9.5x faster pymethod 1.1x faster pyfunction 9.7x faster module_function 1.2x slower load_string_const 9.5x faster load_tuple_const 9.8x faster create_pyobject 9.5x faster create_closure 9.9x faster create_dict 9.8x faster thread_local_read 2.2x slower 

CPython 3.13t results

object_cfunction 9.8x faster cmodule_function 9.6x faster mult_constant 8.8x faster generator 9.5x faster pymethod 9.7x faster pyfunction 9.7x faster module_function 10.0x faster load_string_const 1.8x slower load_tuple_const 9.5x faster create_pyobject 9.5x faster create_closure 9.8x faster create_dict 8.8x faster thread_local_read 2.0x slower 

nogil fork (3.9) results

object_cfunction 10.4x faster cmodule_function 9.2x faster mult_constant 10.0x faster generator 9.0x faster pymethod 9.5x faster pyfunction 10.0x faster module_function 10.0x faster load_string_const 9.8x faster load_tuple_const 9.8x faster create_pyobject 9.6x faster create_closure 9.2x faster create_dict 9.6x faster thread_local_read 9.0x faster 

As mentioned in the PR description, we have known scaling issues related to LOAD_ATTR not using deferred reference counting yet. We also have a scaling issue when reading thread-local data -- we should probably enable deferred reference counting on _thread._local objects.

The 3.13 release avoids the LOAD_ATTR scaling issues due to immortalization. However, we apparently have a bug related to string immortalization (load_string_const is slow) and the thread-local bottleneck is also present.

Note that small variations (e.g. 8.8x vs. 10.4x) are not meaningful.