Support FP8 in radix sort

[gevtushenko]

Description

closes #617
closes #590

This PR:

1. introduces FP8 (__nv_fp8_e4m3 and __nv_fp8_e5m2) support for CUB radix sort
2. fixes block radix sort tests to work correctly in the presence of negative FP values
3. fixes default min / max generator values to be lowest / max
4. fixes c2h generator to correctly handle lowest / max value range

Some of the tests do not work correctly with the lowest / max FP range. These are mostly tests that accumulate values. When significant number of large values (1^38) is accumulated, we push floats to inf values, that complicates comparison. For now, I've reduced the value range for these tests to get them passing. We should think how to handle FP tests going forward.

The PR doesn't update CUB docs to guarantee fp8 support. I suggest the following order:

1. device radix sort test is ported to Catch2
2. device radix sort test covers fp8
3. documentation is updated to guarantee fp8 support

We should also document __nv_bfloat16. For some reason, only __half is documented right now.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[gevtushenko]

/ok to test

[gevtushenko]

/ok to test