gh-95382: Use cache for indentations in the JSON encoder

[serhiy-storchaka]

It is a continuation of #118105 which keeps a cache of newline+indentation and comma+newline+indentation at every level.

For example:

$ ./python -m timeit -s 'import json; a = [[[{"key": "value"}]*10]*10]*10' 'json.dumps(a, indent=2)'
Before: 200 loops, best of 5: 1.79 msec per loop
After:  200 loops, best of 5: 1.25 msec per loop

$ ./python -m timeit -s 'import json; a = [[[list(range(10))]*10]*10]*10' 'json.dumps(a, indent=2)'
Before: 50 loops, best of 5: 5.45 msec per loop
After:  50 loops, best of 5: 4.78 msec per loop

The effect is the strongest if there are many deep narrow trees growing from the root:

$ ./python -m timeit -s 'import json; a = [[[["nested"]]]]*1000' 'json.dumps(a, indent=2)'
100 loops, best of 5: 3.39 msec per loop
200 loops, best of 5: 1.85 msec per loop

And the weakest (up to no difference) if the tree contains few large flat lists or dicts.

• Issue: Remove unused indent_level from Modules/_json.c #95382

[eendebakpt]

When called via the public interface indent_level is always zero. By removing the argument indent_level in the internal implementation this call and some other bits of the code can be a bit simplified.

(this is probably not worth creating a separate PR for, I leave it up to you)

[eendebakpt]

Suggested change

	PyObject *newline_indent = PyList_GET_ITEM(indent_cache, indent_level * 2);
	assert(indent_level * 2 < PyList_GET_SIZE(indent_cache));
	PyObject *newline_indent = PyList_GET_ITEM(indent_cache, indent_level * 2);

[serhiy-storchaka]

The assert above is placed after conditional resizing the cache. It is useful to check that we always get the right cache size, independently from conditions. But here we do not change the cache.

[eendebakpt]

Suggested change

	assert(indent_level * 2 < PyList_GET_SIZE(indent_cache));
	assert(indent_level * 2 - 1 < PyList_GET_SIZE(indent_cache));

[serhiy-storchaka]

indent_level * 2 < PyList_GET_SIZE(indent_cache) is more strong condition.

[eendebakpt]

The PR adds some extra code for creation and usage of the cache, but this simplifies the error handling in the methods using the cache. The performance improvement is a bit larger than I expected which is nice, although it is also fair to say that the performance when writing json files with indent != None is probably not the most important.