wasm2c: Optimize force_read perf and allow embedder to configure policy

[shravanrn]

wasm2c's FORCE_READ_* macros allow spec compliance when using guard pages by preventing elimination of potentially trapping loads, but introduces several low-level performance issues in the generated code. One example is that it prevents load followed by a modification to be prevented from being combined into a “load with modification”​.

Thisoccurs in one the libraries sandboxed by Firefox. The library loads a byte and extends it to 32-bits before processing it --- all in tight loop.

Since wasm2c's load byte corresponds to the following code

u8 result;​
memcpy(&result, &mem[addr], sizeof(u8)); ​
__asm__("" ::"r"(result)); ​
return (u32) result;​

the generated assembly first performs a load with 32-bit load shifted to a byte, sends the byte result to the asm block, and then performs a widening. However a faster option here, is to just read the byte whie sign extending it, and then consume the sign extended result. This only requires moving the last cast prior to the asm. This simple change speeds up this library by 3%. Separately, this PR also allows the embedder to disable force_reads (this is off by default, since it is non conforming), and speeds up the same benchmark by around 5%.