feat(vpto): add A2/A3 MGATHER lowering

[castigli]

Summary

Stacked on hw-native-sys#971.

• lower A2/A3 VPTO GM-to-UB pto.mgather for Coalesce::Row and Coalesce::Elem with Undefined, Clamp, Wrap, and Zero OOB behavior
• support i8/ui8/i16/ui16/i32/ui32/f16/bf16/f32 payloads while requiring exact GM/destination type agreement and signless i32 indices
• use scalar index handling plus 32-byte-aligned per-row MTE2 copies for Row, and direct flat scalar GM/UB access for dense Elem tables
• enforce a fail-closed A2/A3 safe subset: static positive and representable table metadata, row-major none-box UB tiles, no row_plus_one, explicit coalesce, no scratch, exact Row width, and dense contiguous Elem sources
• preserve the reserved EVENT_ID0 scalar-to-MTE2 synchronization contract and reject the deferred GM-to-L1 path
• expose pto.tile.mgather in PTODSL and add NumPy-golden A3 NPU coverage for every Row/Elem and OOB-mode combination
• materialize MGATHER tile handles only for VPTO and clean late identity pointer casts before emission validation

Validation

• hardened Docker ptoas build: passed
• full gather lit directory: 7/7 passed
• A2 and A3 VPTO LLVM compilation: passed for the full supported payload matrix
• broader vpto/ub lit directory: 75/76 passed; the remaining A5 TileLib test is the known Python-bindings-disabled failure (lower_tadd_noop.pto)
• PTODSL vector/cube unit tests: 37 passed, 117 subtests passed
• all eight PTODSL MGATHER kernels compiled through the A3 VPTO toolchain offline
• A3 NPU numerical tests: 8/8 passed, covering Row and Elem with each of Undefined, Zero, Clamp, and Wrap