feat: add wasm SIMD support

This makes wasm version of llhttp around 5-6% faster on real-world payloads. The produced code looks roughly like this: ```c /* Load input */ input = wasm_v128_load(p); /* Find first character that does not match `ranges` */ single = wasm_i8x16_eq(input, wasm_u8x16_const_splat(0x9)); mask = single; single = wasm_v128_and( wasm_i8x16_ge(input, wasm_u8x16_const_splat(' ')), wasm_i8x16_le(input, wasm_u8x16_const_splat('~')) ); mask = wasm_v128_or(mask, single); single = wasm_v128_or( wasm_i8x16_ge(input, wasm_u8x16_const_splat(0x80)), wasm_i8x16_le(input, wasm_u8x16_const_splat(0xff)) ); mask = wasm_v128_or(mask, single); match_len = __builtin_ctz( ~wasm_i8x16_bitmask(mask) ); ``` It is conceptually similar to SSE vectorization that we already support except that we can't multiple comparisons at once and have to check ranges individually. See previous attempt https://github.com/nodejs/llparse/pull/72 Benchmark: https://gist.github.com/indutny/d18d56fe3c7254d888a79eca98f39145 cc @nodejs/llhttp @mcollina