fix(vmi): VMI legality checks, vsstb op, and PTODSL VMI surface enforcement

[Zhendong404]

Summary

Strengthens VMI legality validation and closes PTODSL surface gaps, rebased on current hwsys/main.
Fixes #1058

New op

• feat(vmi): add vmi.vsstb implementation — adds the block-strided store op with a required mask and fixed zero repeat stride, plus lowering guards: pmode="merge" is now rejected by both the verifier (clear diagnostic) and the legacy lowering instead of being silently dropped.

VMI legality checks

• fix(vmi): improve legality check for vmi.vload/vstore — memory element compatibility, block-stride/mask rules, and group divisibility are checked in all mode branches.
• fix(vmi): enhance vcvt conversion attributes — rounding/saturation attribute validation with positive/negative tests.
• fix(vmi): enforce legal lane counts for vreg and mask — unified group verifiers align with the documented domain.
• fix(vmi): add UB-backed memory verification for gather/scatter — vgather/vgatherb/vscatter reject non-UB memory operands at verify time.

PTODSL surface

• fix(ptodsl): enforce VMI lane-count whitelist on vinterpret_cast — the lane count derived by vinterpret_cast is checked against {1,2,4,8,64,128,256} (previously 8xf32 -> si8 silently produced 32 lanes, bypassing the constructor whitelist); the source layout is only preserved when the lane count is unchanged, and a layout-assigned source across a lane-count change is rejected instead of reusing a stale group/slot layout.
• test(ptodsl): improve coverage of vmi dsl tests — inventory/emission coverage for the formal pto.vmi namespace.

Test fixes

• fix(test): fix failed lit testcases — updates tadd_* and soft_postupdate expectations to the deterministic lowering output (constant ordering, set_mask_norm, shared vs duplicated -1 constants) and fixes the vmi_gather_scatter_address_space_invalid RUN line.

Validation

• check-pto: 1538/1539 pass (1 unsupported: vfsim-costmodel feature off)
• check-dsl: 25/25 pass

Segfault fix

Root cause

pto.merge_jit_modules() built each kernel in an independent MLIR Context, then directly cloned operations from later modules into the first module. The cloned operations retained types and attributes owned by the source context. When that context was destroyed, destruction of the merged module accessed freed MLIR objects and could terminate with SIGSEGV in mlir::Operation::~Operation().

Fix

Reparse each source module text in the first module's target Context before cloning its operations. This keeps all operations, types, attributes, and dialect objects in the merged module under one Context lifetime.

Validation

• ptodsl/tests/test_ast_rewrite_example_ir.py: PASS
• check-dsl: 29/29 passed
• check-pto: 1622 tests passed, 0 failures
• Valgrind: no invalid read/use-after-free or pure virtual method called on the merged AST softmax case

[reedhecre]

A3 板测完成（有跳过）

• 触发方式：merged
• 源码提交：9c49c3697de3
• 结果汇总：OK 220 / FAIL 0 / SKIP 3
• 日志：/home/zhongxuan/ptoas-board-monitor/runtime/logs/20260805_012110_merged_pr1087.log
• 结果 TSV：/home/zhongxuan/ptoas-board-monitor/runtime/logs/20260805_012110_merged_pr1087.tsv