BUG: fix CPU feature env diagnostic buffer overruns (#31905) - #31953

Backport of #31905.

PR summary

This is a follow up to gh-30877.

I've had on the back on my mind to verify and check some other issues I saw in the file and finally managed to get to it.

The original PR fixed a buffer overrun in CPU baseline validation. This PR fixes two similar issues in npy__cpu_check_env.

The two issues fixed:

• nonexistent features: an accepted maximum-length environment variable value could write the terminating NUL one byte past the nexist stack buffer
• unsupported dispatched features: repeated unsupported feature names could overflow the notsupp diagnostic buffer, which was sized from NPY_WITH_CPU_DISPATCH rather than from the accepted environment variable length.

Both fixes avoid the trailing-space-and-trim pattern when building the diagnostic strings.

Noting though that the fix for the notsupp buffer increases stack usage in npy__cpu_check_env because the buffer is now bounded by NPY__MAX_VAR_LEN instead of sizeof(NPY_WITH_CPU_DISPATCH) + 1. On my x86_64 laptop, -fstack-usage reports the function growing from 2224 bytes to 3184 bytes, a 960 bytes increase. The function is non-recursive and only runs while parsing CPU feature environment variables during import, however if the size icrease is deemed unacceptable I can change the approach to something like heap allocation, it will require more code changes though.

These tests are mainly sanitizer regressions, in non-sanitized builds, the previous out-of-bounds writes might not crash.

AI Disclosure

I used codex with the gpt 5.5 model. Guided to create the tests, (hence the tests are AI generated) and to review the C code changes, also to verify any incompatibilities (where it verified the stack increase) and to create small reproducers with the same logic for me to actually verify the issues exist.