TileOP: per-dimension static/dynamic B.DIM lowering (SS/SD/DS/DD)

[VV0003]

Summary

Split every lb0/lb1 B.DIM site in template_asm.hpp into four if constexpr branches on ValidCol/ValidRow:

• static dim → B.DIM zero, %c[imm] (assembler compresses to C.B.DIMI)
• dynamic dim → B.DIM %[reg], 0
• lb2 always stays the immediate form (only TPREFETCH's dynamic-GM fallback keeps a runtime register, since physicalCol = rowStride is a runtime value there)

Rebased on top of linx (#79–#81). Includes:

Main changes

• 97 interfaces converted to per-dimension four-branch dispatch (~1436 B.DIM template lines audited, zero illegal forms)
• PTO_MATMUL_HEADER: N/K now immediate-form (resolver is constexpr; the code's own comment admitted "compile-time exact even though passed through GPR"); M stays a register since cooperative group_M is genuinely runtime
• THISTOGRAM fix: the macro's B.DATR %D2, paste produced unmatchable asm (missing space + PadValue); never compiled because no fixture covered it. Template now four-branch parameterized
• TLOAD2/TLOAD4/TSTORE2 lb2 immediate-slot form fixed
• Removed fa-56 legacy dead code: TMAX_T, TSUB_EXP_EXPAND_T, TMUL_ADD_ROWSUM_T, TADD_MUL_EXPAND_T, TMOV_DN2NZ_DYN, DEFINE_TMOV_LAYOUT variants (zero callers; the VPAR raw-opcode spelling does not assemble — superseded by region helpers)
• Fixed pre-existing BLayout::ND2M16/32 → LayoutCvtEnum (also present in Docs/timg2col contract alignment #81's new code)

Verification

• test_v058_engine_contract: 40/40 OK (assertions strengthened: no "ri" masquerade, lb2 immediate invariant, TMATMUL M/N/K forms)
• New ValidShapePerDimSweep.cpp: compile-time sweep of ~65 interface instantiations × four combinations (elementwise, unary, row/col reduce, scalar variants, TCMP/THISTOGRAM/TLOAD_CUBE/TSTORE_CUBE); disassembly form matrix is symmetric (96 static + 96 dynamic per lb0/lb1, lb2 all immediate), zero illegal forms
• Full compile gate: 63/74 fixtures pass; the 11 failures reproduce identically on pristine origin/linx (pre-existing issues: TLOAD "i"-constraint bug, static_assert limits in DynamicShape/TRowMax, Global template missing in ValidShapeImmediate, TImg2Col/TMatmulAllOptions pre-existing match errors)
• B.DIM ordering audit: 100% DIM-before-IOT/IOS (canonical)
• git diff --check clean

[VV0003]

Sweep coverage extended — the first sweep only covered 48 of the 97 four-branch interfaces; a re-audit flagged the remaining 49. Added in 37935d0:

• TPARTADD/MAX/MIN/MUL, TSEL, TFMA, TFILLPAD, TEXTRACT, TINSERT, TEXPANDS, TDEQUANT
• TROWEXPAND* / TROWARGMAX/MIN, TCOLEXPAND* / TCOLARGMAX/MIN (broadcast shape contracts)
• TGATHER / TSCATTER (offset-tile driven dispatch)
• Shared TSTORE ×4 combinations

Now 101 four-combination instantiations, zero compile errors. Disassembly form matrix across the whole fixture:

label	static (C.B.DIMI)	dynamic (B.DIM reg,0)
lb0	170	162
lb1	170	162
lb2	320	0

Symmetric static/dynamic pairs on lb0/lb1, lb2 always immediate, zero illegal forms.

Note on Shared TSTORE: the sweep uses TMOV_L2S_INSERT inline (Shared values must never cross the C++ ABI); passing SharedTile as a noinline function argument triggers the known setRegClass assertion — verified reproducible on the unmodified baseline.

Not sweepable due to pre-existing blockers (identical on baseline): TLOAD2/TLOAD4 (B.DATR ND2NZ alias missing .normal), TIMG2COL/TMATMUL-options match errors, TLOAD plain-form "i"-constraint bug.