GPU trace-gen (part 2): 7 more device-resident fills + preprocessed-table GPU LDE

[ColoCarletti]

Builds on part 1 (#803). Extends the device-resident trace-gen seam to the rest of the per-op tables and moves the remaining tables' LDE onto the GPU.

What this PR does

7 more table fills go device-resident (built directly in VRAM, feeding the LDE/commit with no host upload, same seam as part 1):

EQ, BYTEWISE, MUL, DVRM (ALU) · BRANCH · CPU32 · MEMW (general)

With part 1 this puts the entire per-op arithmetic / control / memory family on GPU (14 tables total).

Preprocessed tables now LDE on GPU. The 6 lookup tables that commit two Merkle trees (BITWISE, DECODE, PAGE, GLOBAL_MEMORY, REGISTER, KECCAK_RC) can't use the device full-column Merkle, so their fill + Merkle stay on host, but the LDE (the dominant cost) now runs on GPU every proof via a new no-Merkle device LDE path.

Net: essentially all main-trace fill + LDE now runs on GPU.

Also: dedupes the per-table dispatchers into one gpu_build_tables<T> helper, and moves the GPU fill-parity tests out of production code into prover/src/tests/gpu_fill_tests.rs.

Correctness

• Per-table byte-parity (multiset for the order-independent dedup ALU tables) between each device fill and its CPU generate_*_trace — 14 tests, all green on RTX 5090.
• e2e prove → verify green, incl. the preprocessed GPU-LDE (proofs with BITWISE/DECODE LDE'd on device still verify; the static-root assert is unchanged and the opening path is untouched).

Performance

• ~4.2% faster on ethrex-10tx (device fills, GPU-on vs CPU-on kill-switch A/B); the preprocessed GPU-LDE is on top of that.

What's left for future PRs

• Trace-gen stages 1–5 on GPU — the memory-model walk is the biggest remaining lever.
• The precompile / continuation tables (COMMIT, HALT, KECCAK×3, ECSM, ECDAS, LOCAL_TO_GLOBAL) — low row-count, low value.