target: add NVIDIA H20 support

[Fucov]

Summary

Add NVIDIA H20 as a first-class GpuTarget and validate the existing Hopper sm_90 AutoMegaKernel path on real H20 hardware.

This is intentionally a minimal hardware-retargeting change: H20 reuses the existing SM90 CUDA implementation and requires no H20-specific scheduler, VM, ABI, or kernel branch.

H20 target

Validated on an NVIDIA H20 96GB system:

• Compute capability: 9.0 (sm_90)
• SMs: 78
• CUDA-visible memory: ~95.1 GiB
• L2 cache: 60 MiB
• Shared memory / SM: 233472 B
• Maximum opt-in shared memory / block: 232448 B
• Nominal HBM bandwidth used by the target: 4000 GB/s

Fresh eval/peak_bandwidth.py measurements:

run 1: 3714.1 GB/s
run 2: 3712.3 GB/s
run 3: 3710.7 GB/s

The target records the median result:

measured_bw_gbs = 3712.3

which is approximately 92.8% of the nominal bandwidth.

Validation

H20 target registration and CUDA execution:

amk doctor                              PASS
SM90 CUDA JIT build                     PASS
ReferenceVM vs eager                    PASS
CUDA megakernel correctness             PASS
multi-token generation vs eager         PASS

The 8-token generation test reports:

matches_eager = true

Correctness-gated single-stream decode baselines:

toy:
  eager   2.0876 ms/token
  AMK     0.3653 ms/token
  speedup 5.72x

llama:
  eager   3.3003 ms/token
  AMK     0.6349 ms/token
  speedup 5.20x

The H20/IR/ABI targeted regression suite passes:

25 passed

including the frozen validator and ABI regression tests.

Existing upstream test status

The full test suite on the H20 host reports:

115 passed, 2 failed

The two failures are:

tests/test_cuda_fp16.py::test_cuda_attention_head_dim_256_static
tests/test_search.py::test_search_rejects_invalid_configs_without_crashing

Both failures were reproduced unchanged on pristine upstream/main at commit a514bbc on the same machine, before applying the H20 target change.

Similarly, ruff check . currently reports five violations in the untouched instructions/reference.py; this PR does not modify that file.

These pre-existing failures are intentionally not changed here in order to keep the H20 hardware-target contribution focused.

Scope

This PR only adds:

• the NVIDIA H20 GpuTarget
• H20 target regression coverage

It intentionally excludes unrelated model-importer, scheduler, attention, search-test, and local benchmark-artifact changes.