fix: stable fgb.ps dot product and N-split for 2048 harts

[CodeDoes]

Fixes the intermittent Code 0 Type 4 mcause=4 crash in the ET Q8_0 multiply path and keeps all 2048 harts busy during generation.

This is a squashed cleanup of the kernel part of fix-q8-dot (what the reviewer asked to keep). The converter was split out to PR #22, and the uberkernel-default-ON change is split out to PR #23.

Changes

1. Fix Q8_0 gather alignment (block_ops.h)

fg32b.ps requires its address to be 32-byte aligned, but block_q8_0.qs starts at offset 2 in the struct, so alignment was unpredictable block-by-block. The old code computed a qs_aligned base + running offsets and split the 32-byte block into up to 4 chunks, patching each address with qs_aligned | ((qs_addr + off) & 31). That required 4 big if/else branches in both q8_dot_tile and q8_dot_compute_x2_aligned.

The fix is to use fgb.ps unconditionally: it gathers 4 bytes by byte offset and has no alignment restriction. Every load becomes fgb.ps f..., f31(%[apN]) with &blk->qs[0/8/16/24]. This removes the alignment math, eliminates the branches, and stops the crash from ever firing.

2. Two-row scale path (block_ops.h, q8_dot_compute_x2_aligned)

The K-block to fp32 scale multiply-accumulate was also streamlined: fbcx.ps now feeds directly into fcvt.ps.f16 in the same cycle, saving a couple of cycles per K-block versus the older extract-then-convert sequence.

3. N-split decode path (mul_mat_Q8_0.c)

Old M-split divided rows across harts: m = hart_id; m < M; m += STRIDE_M. During generation M is typically 1, which meant 2047 of 2048 harts were idle. N-split flips the outer loop so each hart gets a slice of output columns instead: it processes all M rows for its [n_start, n_end) column range. The fast path is enabled when M <= 4 && N > M && N >= 32; the original M-split loop is preserved for large-M workloads in the else branch.

Note: The Q4_0/Q4_K N-split scaffolding from PR #20 also gates on NSPLIT_MIN_N = 10000 so the test frameworks small-N correctness checks still pass. This PR uses the more aggressive M<=4 && N>=32 gate for Q8_0 because the N-split path in Q8_0 duplicates the outer loop rather than wrapping each sub-path.\n\n## Why the diff looks large\nMost of the churn is in block_ops.h: ~300 lines of alignment-branch boilerplate were deleted and replaced by a single concise gather sequence in each function. The actual new instruction logic is smaller and faster; the big headline is simplification, not expansion.\n\n## Review notes\n- Behavior change: the Q8_0 gather method changed from alignment-branched fg32b.ps to unconditional fgb.ps. Correctness hinges on fgb.ps matching or exceeding the gather bandwidth of the old path.\n- Performance: N-split should improve generation throughput when M is small (1-4). Benchmarks on real hardware (DGX-Spark / 2048-hart runtime) are the best way to validate the perf claim.\n- Risk: the duplicate loop nest added for N-split in mul_mat_Q8_0.c is straightforward, but double-check the n_start/n_end edge handling when N is not evenly divisible by 2048.\n\n## Split-out dependencies\n- PR #22 has the SpatialLM Qwen converter (was originally bundled with this kernel work; reviewer asked for conversion/* revert).\n- PR #23 has the GGML_ET_UBERKERNEL default-ON flip, kept out of this PR so reviewers can approve the kernel correctness and the default-behavior change independently.