[NPU][autotuner] config-layer tolerance for Ascend NPU

[yzchang-plus]

Makes Helion's autotuner/config layer robust on Ascend NPU (UB=192KB, triton-ascend) so NPU-incompatible config values degrade silently instead of crashing, and overflowing configs are skipped rather than aborting autotune.

Changes:

• block_id_sequence.py: _normalize discards user values when size==0.
• config_spec.py: downgrade unsupported indexing to NPU ("pointer"); pid_type defaults to "flat"; NPU reduction cap (prod(block)reduction <= UB budget); [None]->default loop; _shrink_for_numel_constraints; imports npu caps.
• logger.py: classify "ub overflow"/"ConvertLinalgRToBinary" as debug.
• local_cache.py: npu_volatile_triton_cache_dir (avoid stale binaries).
• benchmark_provider.py: NPU baseline fallback (halve block_sizes on overflow).
• benchmarking.py / progress_bar.py: do_bench_npu routing + NPU progress.
• _compat.py: safe_clear_cache (clear triton cache for NPU drivers without clear_cache); also holds register_npu_backend / supports_launch_cooperative_grid used by later PRs.
• compile_environment.py: numel constraint substitutes specialized vars and fixed block symbols (so [chunk=256, block_v]-style constraints are created).
• ascend/config.py (+ascend/init.py): npu* env-tunable caps, in subdir.

Problems solved: autotune crashes on NPU-incompatible config keys; UB-overflow configs aborting autotune; stale cached binaries; baseline overflow; numel constraint missing fixed-block dims.

Test Results

NPU (Ascend 910B4)

This PR is the config/autotuner infrastructure layer. It does not run NPU kernels standalone, but is verified as part of the full 4-PR stack (see PR4). On Ascend910B4 with HELION_BACKEND=ascend, the full stack passes add / matmul / softmax / rms_norm / gdn_fwd_h / se_block EXIT=0.

NPU reproduction

HELION_BACKEND=ascend  \
HELION_NPU_MAX_TENSOR_NUMEL=4096 HELION_NPU_UB_BUDGET_ELEMENTS=1024 \
python examples/<op>.py

PR stack

This is PR 1 of a 4-PR stack (merge in order):

1. [NPU][autotuner] config-layer tolerance (this PR)
2. [NPU][codegen] Ascend-safe codegen hooks
3. [NPU][backend] AscendBackend + AST transforms
4. [NPU][runtime] runtime launcher + testing + examples

[yzchang-plus]

Hi @jansel
I noticed this PR hasn't been reviewed yet. I understand you're probably busy — just wanted to check if there's anything I can do on my side to make it easier to review. Thanks for your time!

[yzchang-plus]

Hi @jansel I noticed this PR hasn't been reviewed yet. I understand you're probably busy — just wanted to check if there's anything I can do on my side to make it easier to review. Thanks for your time!

this PR is a follow-up to #3037, this PR makes the autotuner/config layer robust on Ascend NPU — incompatible configs degrade silently instead of crashing autotune, and UB-overflow configs are skipped rather than aborting. Changes are strictly NPU-only guards. Verified end-to-end on Ascend910B4. Happy to walk through or split if needed.

[jansel]

@yzchang-plus how are we testing this? Are all tests now passing if you run:

HELION_BACKEND=ascend pytest test

Looks like you have merge conflicts.