fix(cuda): handle validity in GPU kernels

[0ax1]

All GPU kernels should now handle validity or fall back to the CPU otherwise in cases they don't.

Dynamic dispatch:

• Propagate root array validity through FusedPlan to output.
• Reject Dict with nullable codes and RunEnd with nullable ends (out-of-bounds risk).
• Short-circuit AllInvalid (skip kernel) and empty arrays.

Standalone kernels:

• RunEnd/Zstd: replace unreachable!()/unimplemented!() with vortex_bail!()
  for unsupported nullable cases, enabling graceful CPU fallback.
• ALP: remove stale TODO; patch validity scatter is unnecessary since the
  encoder already strips null positions from the exception list.

[codspeed-hq]

Merging this PR will improve performance by 29.94%

⚡ 2 improved benchmarks
✅ 1120 untouched benchmarks
⏩ 1455 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	take_10k_random	251.2 µs	193.3 µs	+29.94%
⚡	Simulation	take_10k_contiguous	315 µs	257 µs	+22.54%

---

_{Comparing ad/cuda-validity (1aebccf) with develop (ae906c7)}

Footnotes

1. 1455 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[joseph-isaacs]

do you actually do any work in the plan. is that in a different pr

[0ax1]

do you actually do any work in the plan. is that in a different pr

the plan doesn't really have to do sth in addition here other than attaching the validity to the output array. the kernels are fine operating on garbage data. the ones that aren't require non-nullable dict with nullable codes and RunEnd with nullable ends.