Simplify and speed up param packer

[seberg]

This simplifies the param packer and stays in Cython. The original motivation was a small regression found by the benchmarks. But this also speeds up the same cases by a large factor. For many arguments up to 10x.
That many arguments are not too interesting, but also 16 have a 16% speedup at least.

Some further optimizations are certainly plausible and may become more relevant as kernel launch time decreases. I limited it to only a few things:

• Move to Cython (for simplificty)
• Reorder/avoid many isinstance() checks
• Also fixes potentially unaligned memory access. Although, this may just not matter for any platform.

There are a bunch of smaller things I could think of (but didn't do to keep this a more mechanical "move" and I am not sure how hot these are):

• Remove isinstance() checks mostly for exact type checks
• use getattr(type(value), 'getPtr', None) on the type (avoid binding and stricter contract).
• Make the helper a C++ class (may save close to 5% on empty kernel, but added complexity).
• ...

Benchmark

This really changes all except the last two lines (the last two pass an existing buffer or )

Benchmark	main (µs/launch)	This PR (µs/launch)	Speedup
bench_launch_512_doubles	64.78	6.44	10.1×
bench_launch_512_ints	46.05	5.77	8.0×
bench_launch_512_bools	43.54	6.60	6.6×
bench_launch_512_args (all-None types)	26.67	19.16	1.39×
bench_launch_16_args	2.68	2.31	1.16×
bench_launch_512_args_pre_packed	3.17	3.14	1.01×
bench_launch_empty_kernel	1.64	1.62	1.01×

Details

AMD Ryzen Threadripper PRO 7975WX / NVIDIA RTX PRO 6000 Blackwell (sm_120),
driver 580.173.02, CUDA 13.0, CPython 3.14.7 (GIL build), Cython 3.2.9,
g++ -O3. Both sides built through an identical pipeline from a detached-HEAD
worktree and the PR branch. 1500 launches per repetition, min of 5 repetitions,
7 rounds interleaved, median of per-round minima; process pinned to one core on
an otherwise idle machine.

The last two rows are controls: bench_launch_empty_kernel passes
kernelParams=0 and bench_launch_512_args_pre_packed supplies an
already-packed buffer, so neither reaches the packing path and both are
expected to be flat.

[mdboom]

LGTM. My agent found no additional issues.

This is the kind of thing we would normally port to 12.9.x as well, if these patches apply cleanly there. The generated changes will automatically make it there eventually, but they seem independent enough that if the rest weren't ported it would be ok.

[seberg]

but they seem independent enough that if the rest weren't ported it would be ok.

Hmmm, I suspect the generated changes would break without the rest unfortunately. But this should apply cleanly (enough), just means need to follow-up on that.

[seberg]

Created a manual backport in gh-2911, hopefully that is enough to not worry about that!

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.