Add Loop Unrolling and Interleaving as an Optimization

[sppalkia]

These optimizations improve throughput in simple/small loops by amortizing the cost of control flow, providing more opportunities for out of order execution, etc. Currently the following loop in the new single threaded runtime:

|v: vec[i32]| result(for(v, merger[i32,+], |b, i, e| merge(b, e)))

produces the following assembly:

LBB0_2:
	vpaddd	(%rdi), %ymm0, %ymm0
	addq	$32, %rdi
	addq	$-1, %rsi
	jne	LBB0_2

In comparison, an unrolled an interleaved loop would produce something like this (the below is produced by LLVM's auto-vectorizer on a simple C for loop, with interleave count 4:

LBB0_14:
	vpaddd	-480(%rax,%rbx,4), %ymm1, %ymm1
	vpaddd	-448(%rax,%rbx,4), %ymm2, %ymm2
	vpaddd	-416(%rax,%rbx,4), %ymm3, %ymm3
	vpaddd	-384(%rax,%rbx,4), %ymm4, %ymm4
	vpaddd	-352(%rax,%rbx,4), %ymm1, %ymm1
	vpaddd	-320(%rax,%rbx,4), %ymm2, %ymm2
	vpaddd	-288(%rax,%rbx,4), %ymm3, %ymm3
	vpaddd	-256(%rax,%rbx,4), %ymm4, %ymm4
	vpaddd	-224(%rax,%rbx,4), %ymm1, %ymm1
	vpaddd	-192(%rax,%rbx,4), %ymm2, %ymm2
	vpaddd	-160(%rax,%rbx,4), %ymm3, %ymm3
	vpaddd	-128(%rax,%rbx,4), %ymm4, %ymm4
	vpaddd	-96(%rax,%rbx,4), %ymm1, %ymm1
	vpaddd	-64(%rax,%rbx,4), %ymm2, %ymm2
	vpaddd	-32(%rax,%rbx,4), %ymm3, %ymm3
	vpaddd	(%rax,%rbx,4), %ymm4, %ymm4
	subq	$-128, %rbx
	addq	$4, %rcx
	jne	LBB0_14