fix(models): build o_proj row-parallel in the remaining column-parallel families

[gberasmus87]

fix(models): build o_proj row-parallel in the remaining column-parallel families

Seven families pair a column-parallel qkv with LinearOProj (llama, qwen2,
qwen3, qwen3_moe, mistral, gpt_oss, minimax_m2). Three pair one with
LinearReplicated instead: gemma4, qwen3_5_moe and muse_glimmer.

That pairing is wrong under TP>1 in three ways at once. The qkv projection
shards by head -- LinearQKVMerged computes div_even(num_qo_heads, tp_info.size), LinearColParallelMerged shards its output sizes -- so a rank's
attention output is its local head slice rather than the full width. o_proj
therefore has to take the sharded input dim, and the partial sums need an
all-reduce. LinearReplicated keeps the full weight, expects the unsharded
input, and reduces nothing.

It also fails quietly: a missing all-reduce leaves each rank holding a partial
sum that still decodes to fluent-looking text, so a token-level smoke test does
not catch it.

None of the three is reachable today -- all three weight loaders raise
NotImplementedError(... supports TP=1 only) -- so this is a latent trap rather
than a live bug, and the change is a no-op for every current deployment.
LinearOProj degenerates to exactly the previous replicated behaviour at TP=1:
div_even(x, 1) == x, and the all-reduce is skipped when tp_size == 1. It adds
no new constraint from calling get_tp_info() in __init__ either, since each
of these constructors already reaches it one or two lines up through its qkv
class.

The families left alone are correct as they stand: glm5_next, glm_moe_dsa and
minimax_m3 replicate their q/kv projections too, so replicated o_proj is
consistent there, and glm4_moe uses LinearDF11.

Same change #429 makes for qwen4_exp, which @gdevenyi independently validated at
TP=2 (GSM8K 97.00%, unchanged). Doing the remaining three in one pass so the
next family to gain tensor parallelism does not rediscover this.

Verified: LinearOProj and LinearReplicated forwards are bit-identical at TP=1
at each family's real attention dimensions, and each family's test suite shows
the same pass/fail set with and without the change.