PR exo-explore#1978 shipped a head-parallel attention sharding strategy for
DeepSeek V4 (`_shard_v4_attention_heads` interleaved-per-group split of
`wq_b`/`attn_sink`, plus `_AllSumLinear` to reduce wo_b's input across
ranks). On the deployable mlx-community 6-bit checkpoint
(`mlx-community/DeepSeek-V4-Flash-6bit`) the resulting model loads
cleanly and runs to READY but produces saturated-softmax gibberish at
inference: `logprob: 0.0` on every chosen token, empty `top_logprobs`
even at temp=1, output is uniform-distribution BPE fragments
(`'197197197197197197733733Ca733ca…'` on `"The capital of France is"`).

The fault is the head-parallel slicing of V4Attention's LoRA-decomposed
projections. V4Attention's `_grouped_output_projection` manually
reshapes `wo_a.weight/.scales/.biases` based on `n_groups`/`o_lora_rank`/
`head_dim`; the head-parallel `wq_b` slice changes the per-group input
dimensionality that reshape sees, so even though the linear layers all
have valid shapes the math composing them no longer produces correct
activations. Replicated attention sidesteps this entirely.

This PR drops the head-parallel attention sharding and ships a
MoE-only strategy that:

  * replicates attention on every rank (no V4Attention math touched);
  * shards `ffn.shared_experts.{gate,down,up}_proj` and
    `ffn.switch_mlp.{gate,down,up}_proj` (all-to-sharded /
    sharded-to-all on the input/output dim);
  * wraps `layer.ffn` in the existing `ShardedMoEV4` so cross-rank
    reduction happens via `sum_gradients` on input + `all_sum` on the
    MoE output.

Memory profile on Flash (158B-total / 13B-active, 4-bit / mxfp4 experts):

| section          | full sharding | MoE-only   |
| ---------------- | ------------- | ---------- |
| attention/rank   | ~15 GB        | ~30 GB     |
| MoE/rank         | ~65 GB        | ~65 GB     |
| total/rank @ N=2 | ~80 GB        | ~95 GB     |

Both fit on 128 GB nodes. MoE-only also drops one `all_sum` collective
per layer (no wo_b reduction) — single-stream decode is dispatch-bound
on the cluster, so this matters.

Removes ~125 lines of helper code (`_shard_quantized_rows`,
`_AllSumLinear`, `_shard_v4_attention_heads`); adds ~50 lines of
strategy. `ShardedMoEV4` retained — it's the cross-rank reduction
wrapper the new strategy uses.

Bench (mlx-community/DeepSeek-V4-Flash-6bit, 2× M4 Max via Thunderbolt
RDMA, MlxJaccl backend, single 13B-active forward per decode step,
warmup excluded):

| concurrency | aggregate tok/s | per-request tok/s |
| ----------- | --------------- | ----------------- |
| 1           | 34.2            | 34.2              |
| 2           | 44.1            | 22.1              |
| 4           | 52.8 (mean)     | 13.2              |

Notes:

  * Tested with the Blaizzy DSv4 model implementation (`mlx-lm`
    PR exo-explore#1192). Verified MoE-only sharding also loads cleanly with
    rltakashige's variant currently pinned in `pyproject.toml`
    (`leo/deepseek-v4`), but that combination still emits gibberish —
    the bug there is in the model implementation itself
    (Pipeline-mode also fails, so it isn't sharding-related). Tracking
    that separately; this PR is independent of it.
  * No existing DSv4 tests in the repo.

Issue: <link to repro thread / PR comment once filed>