refactor(ptodsl): make signless iN the only integer identity during tracing; demote siN/uiN to wrapper metadata and boundaries

[Zhendong404]

Background

Issue #1405 (fixed by #1408) exposed a structural fragility in PTODSL's scalar lowering: integer values live in two type domains at once.

• The MLIR arith domain only accepts signless iN (per MLIR's design philosophy, signedness is a property of the operation — divsi vs divui — not of the type).
• The authored PTODSL domain carries siN/uiN as value types throughout tracing.

The two domains are bridged by unrealized_conversion_cast strip/restore pairs (_strip_integer_signedness / _restore_integer_signedness in ptodsl/ptodsl/_types.py). This op is designed as a dialect-conversion placeholder ("this type will be legalized away later"), not as an everyday signedness bookkeeping device. Using it this way means every integer value can exist in a split form — a bare signless SSA value paired with an out-of-band signedness annotation — and any tracer path that caches, constant-folds, or otherwise reuses an intermediate SSA value with a stale annotation can emit invalid IR (in #1405: a si32 to i32 cast whose operand was actually an i32 constant).

#1408 fixes the literal-arithmetic instance of this, but the fix is symptomatic: it removes some strip/restore round trips rather than making the value/annotation mismatch structurally impossible. Other reuse paths (constant folding in _tracing/, control flow, sub-kernel boundaries) can still desync the two.

Proposal

Make signless the only integer identity inside tracing:

1. Value layer: during tracing, every integer SSA value is signless iN. siN/uiN never appear as intermediate value types, so _strip/_restore_integer_signedness disappear from hot paths.
2. Metadata layer: the frontend wrappers (_SurfaceValue / RuntimeValue in _surface_values.py) carry signedness as pure Python-side metadata whose only purpose is operator selection (FloorDivSIOp vs DivUIOp, slt vs ult, MaxSI vs MaxUI, ...). Sites that currently reverse-engineer signedness from value.type string prefixes (e.g. _runtime_scalar_ops.py) read the metadata instead.
3. Boundary layer: siN/uiN are still materialized at boundaries — kernel signatures/ABI, VMI storage — with exactly one strip on entry and one restore on exit per value. No cast chains.

After this, an SSA value has exactly one integer identity; signedness no longer lives in the IR at all, so there is nothing for the tracer to desync.

Suggested migration path

1. Land fix(ptodsl): preserve signless arithmetic operand types #1408 (after fixing its unsigned-semantics regression) as the stopgap.
2. Add a signedness metadata field to the runtime-value wrappers; make _runtime_scalar_ops.py prefer metadata and fall back to the type. This step is independently testable.
3. Remove _strip/_restore from hot paths module by module (_scalar_adaptation.py, _scalar_coercion.py, _runtime_index_ops.py, control flow, sub-kernels), keeping them only at ABI/storage boundaries (consider renaming to _adapt_abi_signedness so misuse is visible in review).
4. Regression coverage: signed and unsigned runtime scalars mixed with literals across add/sub/mul/floordiv/mod/compare/min/max, including values >= 2^31 for the unsigned cases.

Non-goals

• Changing the public DSL surface: pto.si32/pto.ui32 declarations and their numeric semantics stay as they are.
• Changing kernel ABI types.

Refs: #1405, #1408