[RVV] f16 softmax for rvv

[ken-unger]

Add rvv kernel for f16-raddstoreexpminusmax (u4v productized)

Tested on qemu & bpi-f3.

f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u1v/N:7680/real_time       18150 ns        18098 ns        38459 bytes=1.69258G/s cpufreq=1.6G elements=423.145M/s
f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u1v/N:65280/real_time     136772 ns       136277 ns         5118 bytes=1.90916G/s cpufreq=1.6G elements=477.29M/s
f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u2v/N:7680/real_time       11620 ns        11601 ns        60322 bytes=2.64369G/s cpufreq=1.6G elements=660.924M/s
f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u2v/N:65280/real_time      91274 ns        90820 ns         7689 bytes=2.86085G/s cpufreq=1.6G elements=715.212M/s
f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u4v/N:7680/real_time        9968 ns         9956 ns        70297 bytes=3.082G/s cpufreq=1.6G elements=770.5M/s
f16_raddstoreexpminusmax/rvvfp16arith_rr2_p2_u4v/N:65280/real_time      89592 ns        89278 ns         7082 bytes=2.91456G/s cpufreq=1.6G elements=728.639M/s

Note that the change to avx_microkernels cmake/bzl (avx-rr2-p5-nr2-u16) was via tools/update-microkernels.py which fixed a previous unrelated commit (Respect numerical consistency flag in sigmoid configs)