ET backend: N-split infrastructure for Q4_0, Q4_K, Q8_0 MUL_MAT kernels

[CodeDoes]

Adds an N-split sharding path to the ET Q4_0, Q4_K, and Q8_0 MUL_MAT kernels so that all 32 shires are utilized during generation when the output-column dimension (N) is large.

This is a squashed cleanup of PR #17's infrastructure (previously 4 commits: bffd006 + 3 syntax fixups). The guard threshold is intentionally set high so that tests with small N keep the old “every shire computes full output” behavior, which is what the test infra validates.

How N-split works

Old M-split: for (m = hart_id; m < M; m += STRIDE_M). During autoregressive generation M is typically 1, so 31 of 32 harts are idle and every hart still loads the full weight matrix from DRAM.

New N-split: each shire (64 harts = 1 shire) computes a contiguous [n_lo, n_hi) column slice of the output. The loop body for every path in every kernel is now bounded by n_lo .. n_hi instead of 0 .. N.

int64_t n_lo = 0;
int64_t n_hi = N;
if (N >= NSPLIT_MIN_N) {
    const int64_t shire_id = hart_id >> 6;        // 0..31
    const int64_t N_per_shire = (N + 31) / 32;
    n_lo = shire_id * N_per_shire;
    n_hi = min(n_lo + N_per_shire, N);
}

Files changed

• mul_mat_Q4_0.c — add NSPLIT_MIN_N gate + n_lo/n_hi before every for (n = 0; n < N; n++) loop (K-split path, K-split-grouped path, tile-outer scalar path, simple path). Net +35 / −15.
• mul_mat_Q4_K.c — identical scaffold, identical gate. Net +35 / −15.
• mul_mat_Q8_0.c — identical scaffold, identical gate. Note the Q8_0 file here is still the simple n_lo/n_hi slice added on top of the existing M-split; the separate q8_dot_tile / q8_dot_compute_x2_aligned gather rewrite (which eliminates the alignment crash) lives in PR fix: stable fgb.ps dot product and N-split for 2048 harts #21. Net +33 / −21.

Why NSPLIT_MIN_N = 10000 (Q4_0/Q4_K) / 100000 (Q8_0)

At very small N, every shire hits the same atomic_store_f32 destination and the shire distribution adds 32× more cache-line invalidations than a single hart doing full N. Setting the threshold high keeps the fast path disabled until N is large enough that the bandwidth win outweighs the atomics cost. The threshold is deliberately conservative; tune it after real hardware benchmarks.

Review flags

• The body of the if (N >= NSPLIT_MIN_N) branches in Q8_0 does not yet duplicate the single-hart path — it still relies on the existing M-split distribution, so correctness for large-M workloads is unchanged.
• Note the Q8_0 threshold is 100000 here vs 10000 in the Q4_0/Q4_K files; PR fix: stable fgb.ps dot product and N-split for 2048 harts #21's production path later changes the Q8_0 activation condition to M <= 4 && N >= 32. Expect to reconcile the two thresholds when merging.