Reduce the computational complexity of formatting

[cuviper]

This only matters for very large values, more than about 1000 digits, but it does make a big difference for those.

For the included benchmarks, I got these base-10 results before:

test to_str_radix_10      ... bench:       2,095 ns/iter (+/- 20)
test to_str_radix_10_2    ... bench:     150,702 ns/iter (+/- 1,231)

The _2 variant is over 10,000 bits. With the new code, I get these results:

test to_str_radix_10      ... bench:       2,065 ns/iter (+/- 22)
test to_str_radix_10_2    ... bench:      44,130 ns/iter (+/- 231)

[cuviper]

Less rigorously, I swept over a series of large values to help determine the threshold for enabling this new code.

Before:

2^2^10-1 has 309 digits, 3.14µs elapsed
2^2^11-1 has 617 digits, 7.14µs elapsed
2^2^12-1 has 1234 digits, 26.64µs elapsed
2^2^13-1 has 2467 digits, 99.961µs elapsed
2^2^14-1 has 4933 digits, 390.405µs elapsed
2^2^15-1 has 9865 digits, 1.560669ms elapsed
2^2^16-1 has 19729 digits, 6.168387ms elapsed
2^2^17-1 has 39457 digits, 24.690763ms elapsed
2^2^18-1 has 78914 digits, 98.170569ms elapsed
2^2^19-1 has 157827 digits, 395.343446ms elapsed
2^2^20-1 has 315653 digits, 1.572607067s elapsed

After:

2^2^10-1 has 309 digits, 2.98µs elapsed
2^2^11-1 has 617 digits, 7.03µs elapsed
2^2^12-1 has 1234 digits, 16.53µs elapsed
2^2^13-1 has 2467 digits, 39.081µs elapsed
2^2^14-1 has 4933 digits, 105.181µs elapsed
2^2^15-1 has 9865 digits, 323.034µs elapsed
2^2^16-1 has 19729 digits, 1.122664ms elapsed
2^2^17-1 has 39457 digits, 3.710655ms elapsed
2^2^18-1 has 78914 digits, 13.514685ms elapsed
2^2^19-1 has 157827 digits, 48.508852ms elapsed
2^2^20-1 has 315653 digits, 187.755271ms elapsed

I honestly expect there's still room for improvement, but this is so much better that I think it's worth it already.

[cuviper]

bors r+

[bors]

Build succeeded:

• Format
• No Std
• Test (1.31.0)
• Test (1.34.0)
• Test (1.36.0)
• Test (1.40.0)
• Test (1.46.0)
• Test (beta)
• Test (i686)
• Test (nightly)
• Test (stable)