Speed up use of PBKDF2 when building in Debug configuration

[gwynne]

MongoKitten users (including those using the Fluent 4 driver built on it) sometimes experience significant connection delays during development due to the compiler not optimizing the PBKDF2 implementation. With a little help from Instruments, these delays can be massively reduced. The follow shows a timing comparison between the original code and that provided in this PR:

Comparisons

10,000 iterations of each hash are repeated 10x, and the total times averaged. Each test is run in Debug and Release configurations, against the original and revised implementations. The first number is the runtime average of the original code; the second number is the runtime average of the new code. The third number expresses how much of the original implementation's runtime is used by the revised one.

macOS (Xcode 15 beta 5)

Config	MD5	SHA-1	SHA-256
debug	5.736s / 0.256s (4.6%)	10.268s / 0.491s (4.8%)	13.799s / 0.483s (3.5%)
release	0.066s / 0.064s (97.0%)	0.099s / 0.089s (90.0%)	0.126s / 0.116s (92.1%)

Linux (swift:5.8-jammy)

Config	MD5	SHA-1	SHA-256
debug	1.711s / 0.297s (17.4%)	2.088s / 0.634s (30.2%)	3.135s / 0.631s (20.1%)
release	0.119s / 0.067s (56.3%)	0.083s / 0.082s (98.8%)	0.100s / 0.095s (95.0%)

Linux (5.9 nightly - swift-5.9-DEVELOPMENT-SNAPSHOT-2023-07-25-a)

Config	MD5	SHA-1	SHA-256
debug	1.379s / 0.290s (21.0%)	1.435s / 0.565s (39.4%)	1.625s / 0.715s (44.0%)
release	0.121s / 0.070s (57.9%)	0.083s / 0.085s (102%)	0.130s / 0.124s (95.4%)

[gwynne]

For comparison, here's a quick set of performance numbers for using the SwiftCrypto Insecure.MD5, Insecure.SHA1, and SHA256 implementations:

macOS (Xcode 15 beta 5)

Config	MD5	SHA-1	SHA-256
debug	0.141s	0.119s	0.120s
release	0.087s	0.067s	0.067s

Linux (swift:5.8-jammy)

Config	MD5	SHA-1	SHA-256
debug	0.340s	0.237s	0.258s
release	0.076s	0.063s	0.065s

Linux (5.9 nightly - swift-5.9-DEVELOPMENT-SNAPSHOT-2023-07-25-a)

Config	MD5	SHA-1	SHA-256
debug	0.352s	0.248s	0.280s
release	0.074s	0.062s	0.064s

Conclusions:

• SwiftCrypto's MD5 implementation is oddly slow. (This might just be a side effect of MD5 having relatively little internal state, or possibly the M1 just handles SHA-1 and 2 better 🤷‍♀️)
• SwiftCrypto's SHA-1 implementation is much faster than us.
• SwiftCrypto's SHA-2 implementation is massively faster than us.

Might be the numbers would change all around on an Intel machine, though.