mega_moe: add a mori HIP dispatch backend

[jhchouuu]

Makes MegaMoEGfx1250's first stage swappable: dispatch_backend="mori" (or
MEGA_DISPATCH=mori) runs mori's HIP/JIT gfx1250 TDM dispatch instead of this
package's FlyDSL one. Nothing else in the pipeline changes — gemm1, the gemm2
P2P scatter epilogue and the fused combine are untouched, and the default is
unchanged, so not setting anything is byte-for-byte today's path.

Why it composes

The two dispatches leave identical state in the arena:

field	encoding, both sides
disp_out	rows at slot * hidden
out_idx / out_wts	[slot * topk + k], full topk row
flat dest map	dest_pe * max_recv + slot, null npes * max_recv
recv_to_src_token	src_pe * max_tok_per_rank + src_tok
recv_num / tok_off	same signal/ack handshake (wait 0, store count+1, reader restores 0)

The reverse map is the one field the fused path actually consumes — the GEMM
host pass decodes it to build ep_rowmap. mori's dispatch also never touches
cross_device_barrier (its cross-rank rendezvous is the recv_num
handshake), so the fused combine's phase counter is undisturbed.

mori's plan layer takes an arena by duck typing (.handle / .offset), so
SymmetricArena is handed over as-is: no new region, no extra device memory.

Two things that do differ, handled here

• Geometry comes from mori's own tuning table, not _select_dispatch_config.
  That table asks for 32 warps above 256 tokens, and mori's dispatch stages a
  hidden-dim tile per warp in dynamic LDS — 32 * 7168 * 2 = 458 KB against a
  320 KB budget. EpCfgIsValid does not check LDS, so it would fail at launch
  rather than when the plan is built.
• total_recv is zeroed on the host. This package's dispatch clears it in
  its own Phase 2; mori's only accumulates, and the fused combine never resets it.

One visible consequence: the recv slot a token lands in changes, because mori's
gfx1250 dispatch reserves a block's slots with one atomic and hands them out
block-local. Nothing indexes by slot order, but a test diffing arena contents
slot-by-slot against the FlyDSL dispatch will see it.

Measured

4x gfx1250, EP4 hidden 7168 topk 6, --combine scatter_fused, against
test_mega_moe.py's fp32 reference:

config	FlyDSL	mori	logits_diff
2 layers, 256 tok/rank	945.9 us/layer	925.8	0.002174 (both)
4 layers, 4096 tok/rank	2678.0 us/layer	2622.2	0.004344 (both)

The diffs are identical because dispatch only moves data — the deviation from
the reference should not move, and it does not. --profile_table 1 confirms the
geometry that actually launches is mori's: ..._64x8 at 256 tokens,
..._64x16 at 4096.

Requires

A mori built with JIT v2 (ROCm/mori#548 or later) and its libmori_ops_v2.so.
Import failure is reported with which half is missing. Selecting the backend
costs ~5 s at construction (three geometries, one JIT compile each), before any
cudagraph capture.