fix(rocm): gather QMV large-batch launch and 64-bit output offsets #2620

Problem

Follow-up to upstream ml-explore/mlx#4738 (53c18705, CUDA gather-QMM int64 strides). The ROCm default gather paths already use int64 expert offsets (quantized/qmm.hip 4759-4766, 5010-5018, 5206-5224, 5675-5690, 5779-5790, 5877-5891, 5998-6001; quantized/qmv_tiled_kernel.hip 436-443), but 32-bit sites and a possible launch limit remain. Paths are under src/lib/mlxcelverse/overlays/rocm/mlx/backend/rocm/.

32-bit sites

• quantized/qmv_tiled_kernel.hip:342-345: gather_qmv_tiled_kernel computes rhs_idx * N * w_stride and lhs_idx * M * K with uint32_t indices. It is opt-in through MLX_ROCM_GATHER_QMV_USE_TILED (qmm.hip:6447-6449).
• int output index batch * M * N + row * N + col at qmv_tiled_kernel.hip:394, 501 and qmm.hip:5870, 5929 (gather_qmv_kernel). qmm.hip:5608, 5714, 5819 already use int64 and are the pattern to copy.
• gemms/gemv.hip:272: out + indices_idx * rows with int indices_idx and int rows.
• quantized/qmv_kernel.hip repeats the pattern but is dead: not in CMakeLists.txt and included nowhere.

Possible launch limit (inferred, not verified)

The gather kernels put batch B = tokens * top_k in grid z (qmm.hip:5170, 5643, 5749, 5844; fast_grid at 6214; tiled gt_grid at 6500). The header of tests/rocm_hip_grid_clamp.rs:18-26 states y and z both top out at 65535 blocks, but only the y cap was observed failing (:48-52). If z has the cap, B > 65535 fails to launch; a 2048-token prefill chunk (src/lib/mlxcel-core/src/prefill_plan.rs:61) at top_k 32 reaches it.

Acceptance Criteria

Implement all of it in one PR with one verification pass at the end.

• Decisive test first in tests/rocm_hip_grid_clamp.rs: unsorted gather_qmm, bf16, affine 4-bit, group 64, E=128, M=1, N=K=64, B=65,537, against CPU. Record in the PR whether the z cap is real
• If it is, every gather QMV kernel indexing batch by blockIdx.z folds the batch into a grid capped at 65535 and loops over batch with a stride (not needed: the z cap was ruled out on gfx1151, see fix(rocm): widen gather QMV output offsets to 64 bits #2624)
• Listed offsets widened to int64; an #[ignore] test (f32 activations, N=65,536, K=64, B=32,769, output past 2^31 elements) checks the last row against CPU
• Optional: delete quantized/qmv_kernel.hip and its mention in qdequant.hpp:2
• LOCAL_FIXES.md entry (next free number) and make verify-rocm-overlay green

Verification

bash scripts/rocm_gpu_guard.sh --hold -- cargo test --profile test-fast --features rocm --test rocm_hip_grid_clamp -- --test-threads=1 --include-ignored
make verify-rocm-overlay

This edits the same files as #2617 (native NVFP4) and should land before it.