Add basic HVX reduce kernels.

[copybara-service]

Add basic HVX reduce kernels.

Kernels that should be reasonably good:

• min, max, min_max for all types
• sum, sum_squared for int8 and uint8 for k1 > 1

Kernels that are not good and need work:

• sum, sum_squared for int8 and uint8 for k1 = 1. These are currently naively implemented with conversions, and wide arithmetic (instead of widening arithmetic).
• In general k1 = 1 is not good because we unroll the accumulator by 2x/4x, so we can load whole vectors, which makes the accumulators really large (e.g. 128). This means that we're very likely to hit tail case code.

Example inner loop (k1 > 1 uint8 sum, sum_squared is almost identical):

.LBB30_116:                             // %while.body14.i
                                        //   Parent Loop BB30_110 Depth=1
                                        //     Parent Loop BB30_112 Depth=2
                                        //       Parent Loop BB30_114 Depth=3
                                        // =>      This Inner Loop Header: Depth=4
        {
                v27 = vmemu(r5++#1)
        }
        {
                v28 = vmemu(r6++#1)
        }
        {
                v10.w += vrmpy(v27.ub,r9.b)
                v23 = vmemu(r0++#1)
        }
        {
                v9.w += vrmpy(v28.ub,r9.b)
                v15 = vmemu(r7++#1)
        }
        {
                v5.w += vrmpy(v23.ub,r9.b)
        }
        {
                v6.w += vrmpy(v15.ub,r9.b)
                r3 = add(r3,#-128)
        }
        {
                p3 = cmp.gtu(r3,#127)
                if (p3.new) jump:t .LBB30_116
        }