TileOP: eliminate scalar stack round-trips, fix SD-branch dim binding (SuperNPUBench#112)

[VV0003]

Fixes the TileOP-side of SuperNPUBench#112 (plus a related binding bug found during verification).

Type A: keep_acc() zero-descriptor dead code

Analysis on current linx: already fixed by #71 — no sdi/ldi is emitted before B.FPATR for parameter-free options (verified by repro). This PR removes the residual hazard: the volatile descriptor references now live inside the if constexpr (IorMode != 0) guard so parameter-free options cannot resurrect the dead materialisation under any future optimisation level.

Type B: scalar constant stack round-trips (swi/lwi)

Confirmed and fixed. Root cause: the anti-fold idiom volatile T sv = s; (78 occurrences across TEXPANDS/TMULS/TADDS/TCMP/TQUANT/TDEQUANT/TCMP/TPREFETCH/...) forces the scalar through the stack:

lwi.u [const pool] -> x0
swi   x0, [sp, 204]      # redundant
lwi   [sp, 204], -> a4    # redundant

Replaced with an empty asm register tie asm("" : "+r"(sv)) — the value stays in a GPR (anti-fold preserved: B.IOR binds a real register, verified with the 0.0f worst case) and constants load directly from the pool into the asm input register:

lwi.u [const pool] -> a1     # direct

Bonus fix: SD-branch dim binding mixup

While auditing whether any static value still reaches B.DIM reg, 0, found three branches (static ValidCol / dynamic ValidRow: TMUL, TROWSUM, TROWEXPANDMUL) that bound [valid_col] "r"(GetValidCol()) into lb1 — the wrong dimension, and a compile-time constant into a register operand. This is exactly what produced "register fed by an immediate then B.DIM reg, 0". Fixed to bind GetValidRow(). After the fix: all 324 register-form B.DIM operands in the per-dimension sweep trace to genuine runtime defs (loads/parameters), zero immediate-fed.

Verification

• Issue repro (TMATMUL+keep_acc / TEXPANDS constant / TMULS constant): zero stack round-trips, zero dead descriptors
• B.IOR anti-fold preserved (constant 0.0 still binds [a1] via pool load, never [zero])
• Full gate 64/74 — the 11 failures are identical on the unmodified baseline
• unittest 40/40

Note: the reported ELF was built with the 0904 toolchain (1ae4ee39); the header-side fixes here apply to any toolchain version. Backend-side constant rematerialisation (c.movr direct move where a stack slot was used) is covered by the same header change — no backend edit needed.