lang: Update dispatch function to support dynamic discriminators

[acheroncrypto]

Problem

The generated dispatch function gets the instruction discriminator by taking the first 8 bytes of the instruction data:

anchor/lang/syn/src/codegen/program/dispatch.rs

Lines 57 to 65 in e5bed20

	// Split the instruction data into the first 8 byte method
	// identifier (sighash) and the serialized instruction data.
	let mut ix_data: &[u8] = data;
	let sighash: [u8; 8] = {
	let mut sighash: [u8; 8] = [0; 8];
	sighash.copy_from_slice(&ix_data[..8]);
	ix_data = &ix_data[8..];
	sighash
	};

This obviously doesn't work with discriminators that are not exactly 8 bytes.

Summary of changes

Dispatch instructions based on whether the given data starts with an instruction's DISCRIMINATOR constant.

Bench

We save a little bit of CU initially due to not initializing an intermediary sighash variable. However, comparing via starts_with as opposed to matching exact bytes has a slight overhead of ~3 CU per instruction. The result is:

• If the program has less than 20 ixs, this results in a slight performance gain
• Otherwise, this results in extra ~3 CU per each instruction, e.g. if a program has 50 instructions, the last instruction will use ~100 more CU compared to before this PR

See the compute units benchmarks to compare with real values as instruction count increases (bench program has 87 instructions).

The difference shouldn't be noticable in almost all cases. However, if people notice the regression with higher instruction count programs, we can consider adding a feature flag that uses the old dispatch codegen.

---

Note: This PR is part of a greater effort explained in #3097.

[vercel]

@acheroncrypto is attempting to deploy a commit to the coral-xyz Team on Vercel.

A member of the Team first needs to authorize it.