Fixes performance of convertTo with multiplication on RISC-V RVV 0.7.1 (tested on LicheePi 4A with Xuantie 2.8.0 toolchain):

• use madd/fmadd (a = a * b + c) instead of macc/fmacc (c = a * b + c) - it helps to use v-registers efficiently
• modified v_load/v_store - for some reason 8-bit load/store instructions were used - avoid type transformations and extra instructions being added

Some tests slowed down (up to 5 times), but about a quarter of core tests improved performance up to 10 times. Similar picture with imgproc, but about a half of tests improved up to 6 times. I'm going to improve other v_load intrinsics and investigate ways to fix slowed cases.