use undici.request for httpClient

[pushrax]

according to the undici docs this method is much faster than fetch https://github.com/nodejs/undici/tree/main#benchmarks

my preliminary performance testing agrees with this, notably better combined body_read_time+deserialize_time when the process is under load

this is passing tests and should work, but could use:

• [x] test that gunzip actually works (the test suite responses are too small to trigger it)
• [ ] stream gunzip? edit: decided not to
  • advantages of streamed gunzip:
    • in theory it's better to pipeline the cpu work with the io
  • advantages of oneshot gunzip:
    • can time the body read and gunzip separately
    • maybe drive i/o faster to buffer the whole response sooner
    • fewer context switches to C
• [x] separate metric for gunzip / json?
• [x] reduce lint errors

[sirupsen]

I guess for non-streamed gunzip, it's intuition, so a benchmark will always be better if it's not a huge PITA.

[pushrax]

• sticking with zlib, should be more efficient and worst case with pako is worse
• sticking with oneshot gunzip for now since measurement is clear, can optimize it later
• keeping the additional metric load to the most important ones for now