riscv64/mc: Add bidir functions

This code compromises between the performance of a dedicated kernel per
VLEN/width pair, and the flexibility of a fully VLEN-dynamic loop, by
using a single special case for w=4, and subdividing the rest into the
unrolled four line fast path, and the general-purpose slow path (for
large width on small VLEN).

Kendryte K230

avg_w4_8bpc_c:          346.8 ( 1.00x)
avg_w4_8bpc_rvv:         50.3 ( 6.90x)
avg_w8_8bpc_c:         1054.9 ( 1.00x)
avg_w8_8bpc_rvv:        139.1 ( 7.58x)
avg_w16_8bpc_c:        3396.3 ( 1.00x)
avg_w16_8bpc_rvv:       350.6 ( 9.69x)
avg_w32_8bpc_c:       13734.3 ( 1.00x)
avg_w32_8bpc_rvv:      1226.3 (11.20x)
avg_w64_8bpc_c:       33260.9 ( 1.00x)
avg_w64_8bpc_rvv:      3869.4 ( 8.60x)
avg_w128_8bpc_c:      83441.3 ( 1.00x)
avg_w128_8bpc_rvv:     9765.1 ( 8.54x)

w_avg_w4_8bpc_c:        444.3 ( 1.00x)
w_avg_w4_8bpc_rvv:       75.8 ( 5.86x)
w_avg_w8_8bpc_c:       1365.6 ( 1.00x)
w_avg_w8_8bpc_rvv:      208.8 ( 6.54x)
w_avg_w16_8bpc_c:      4420.8 ( 1.00x)
w_avg_w16_8bpc_rvv:     570.7 ( 7.75x)
w_avg_w32_8bpc_c:     18010.9 ( 1.00x)
w_avg_w32_8bpc_rvv:    2074.4 ( 8.68x)
w_avg_w64_8bpc_c:     43050.4 ( 1.00x)
w_avg_w64_8bpc_rvv:    5799.5 ( 7.42x)
w_avg_w128_8bpc_c:   107153.6 ( 1.00x)
w_avg_w128_8bpc_rvv:  14272.0 ( 7.51x)

mask_w4_8bpc_c:        497.6 ( 1.00x)
mask_w4_8bpc_rvv:       88.5 ( 5.63x)
mask_w8_8bpc_c:       1528.5 ( 1.00x)
mask_w8_8bpc_rvv:      253.1 ( 6.04x)
mask_w16_8bpc_c:      4953.8 ( 1.00x)
mask_w16_8bpc_rvv:     679.0 ( 7.30x)
mask_w32_8bpc_c:     20298.3 ( 1.00x)
mask_w32_8bpc_rvv:    3012.9 ( 6.74x)
mask_w64_8bpc_c:     49718.8 ( 1.00x)
mask_w64_8bpc_rvv:    7291.7 ( 6.82x)
mask_w128_8bpc_c:   126740.3 ( 1.00x)
mask_w128_8bpc_rvv:  18351.1 ( 6.91x)