bpo-37295: Speed up math.comb(n, k) for 0 <= k <= n <= 67

[mdickinson]

This PR speeds up math.comb(n, k) for very small k and n.

The existing tests already include exhaustive testing of correctness of math.comb(n, k) for 0 <= k <= n <= 100, so I didn't add any more unit tests.

Some sample timings on my machine (comparing commit fd3a4bb on this branch with commit 2e3e0d2 on the main branch):

./python.exe -m timeit -s "from math import comb;n, k = 1, 1" "comb(n, k)": difference not significant

• main: 5000000 loops, best of 5: 35.5 nsec per loop
• here: 10000000 loops, best of 5: 34.2 nsec per loop

./python.exe -m timeit -s "from math import comb;n, k = 5, 3" "comb(n, k)": ~23% faster (but may not be significant)

• main: 5000000 loops, best of 5: 38.6 nsec per loop
• here: 10000000 loops, best of 5: 31.5 nsec per loop

./python.exe -m timeit -s "from math import comb;n, k = 67, 25" "comb(n, k)": ~9.7x faster

• main: 500000 loops, best of 5: 447 nsec per loop
• here: 5000000 loops, best of 5: 46.1 nsec per loop

https://bugs.python.org/issue37295

[tim-one]

Very nice and very clear. Ship it 😃