Speed up randint array-bounds via chunking Lemire

[vlad-perevezentsev]

This PR implements the word-generation speedup suggested for randint array-like bounds in the #168 (comment)

It changes irk_rand_bounded_broadcast to:

• consume MKL words in cache-sized chunks instead of one full-length pass, avoiding a memory round-trip over the whole array;
• pick the Lemire rejection test by range width: lo < s is free for narrow ranges but mispredicts for wide ones, so its hits are counted on the first chunk and the remaining chunks switch to the branchless test with the threshold memoized per range.

Performance (Intel(R) Xeon(R) Platinum 8480+)

(low=0, high=1e9):

Size (N)	dtype	Before	After	Speedup
5,000,000	int32	52.00 ms	31.44 ms	1.65x
5,000,000	int64	48.82 ms	35.53 ms	1.37x
50,000,000	int32	642.91 ms	428.80 ms	1.50x
50,000,000	int64	602.30 ms	466.51 ms	1.29x

[ndgrigorian]

One idea is to move this loop out, and track n_pending outside of the chunk loop (i.e., before the for loop) then retry all of the rejects at the end. We'd have to set the idx allocation to len instead of chunk_cap too.