[Hyphen-05]

MilkdropShader::TranspileHLSLShader() removes sampler and texsize declarations with two
while (std::regex_search(...)) loops. Each pass constructs the regex again and searches the whole shader from the
start. On a Moto G Play (2023), 32-bit ARM with libc++, the two loops take about 7.5 ms per shader, so 15 ms on every
preset load. That is part of the stutter on preset switches reported in #913.

The two regexes are now built once when the library loads. That alone makes no measurable difference, because
constructing them only takes about 25 µs. The time is spent in regex_search trying the pattern at every position.
Both patterns start with a fixed word, so the new Utils::EraseRegexMatches() finds that word with
std::string::find() and only tries the regex there, using match_continuous. After each erase it searches again
from the start, like the old loop, because an erase can create a new match further up.

29 shaders from Cream of the Crop presets, per shader, on the same phone:

	median	90th percentile
before	7.52 ms	10.50 ms
regexes built once only	7.55 ms	10.49 ms
this PR	0.19 ms	0.25 ms

The output is identical to the old loop for all 29 shaders, and for 300,000 random strings made from sampler, 2D,
3D, float4, texsize_, ( and whitespace.

Added unit tests for EraseRegexMatches(), including a case where an erase creates a new match earlier in the string.

[kblaschke]

LGTM

Thanks for the contribution!