RISC-V: register-block the ZVL256B GEMV-N kernel

[6eanut]

Summary

Process four consecutive columns per pass over y in the ZVL256B GEMV-N kernel
for unit y stride.

The kernel previously walked one column of A at a time. For each column it
loaded y, loaded the column, issued one vfmacc, stored y back, and then
repeated the whole sequence for the next column. Every column therefore paid a
y load and a y store round-trip, and because VL was recomputed from the
remaining row count on each pass, the vsetvli and the pointer bumps also sat
inside the hot loop.

GEMV_N_BLOCK (4) consecutive columns are now accumulated into a single y
slice that stays resident in vector registers, so the y round-trip is paid
once per block instead of once per column, and the main loop's VL is pinned to
the hardware maximum, which lets the compiler hoist the vsetvli and
strength-reduce the pointer arithmetic out of the loop.

The columns are still consumed in ascending j and the four contributions are
folded into the accumulator in that same order, so each y[i] sees the same
sequence of addends as the original column-at-a-time loop; the intermediate y
vector merely stays in a register instead of being stored and reloaded. The row
remainder of each block keeps a variable VL, and the columns left over after the
last full block are finished by the two-column and single-column loops. The
result is bit-identical.

The non-unit y stride path is untouched.

This backs both SGEMV_N and DGEMV_N.

Performance

benchmark/sgemv.goto and benchmark/dgemv.goto, one X100 core at 2.2 GHz,
single thread, warm-up once and the median of three runs per size:

Benchmark	Size	Baseline	Patched	Change
sgemv	256	5885.21	5817.72	-1.15%
sgemv	512	4848.66	6303.53	+30.01%
sgemv	1024	4584.54	5206.73	+13.57%
dgemv	256	2410.66	3158.90	+31.04%
dgemv	512	3670.95	4483.65	+22.14%
dgemv	1024	1899.48	1945.03	+2.40%

Testing

• make tests returns 0: 125/125 utests, 1473/1473 extension tests, no new
  CBLAS failure.
• Differential test under QEMU with guard pages around every buffer: 21375
  single-precision, 21375 double-precision and 21375 non-unit-y-stride cases,
  covering m = 1..129, n = 1..33, several leading dimensions and x
  increments, and alpha values including 0, produce output bit-identical to the
  unpatched kernel, with no out-of-bounds access. Both builds were also checked
  against a double-precision reference.
• The inc_y != 1 branch was diffed against the baseline and is byte-identical.
• An end-to-end call through the public sgemv_ entry point on both the
  baseline and patched libraries produces identical results.