fix: Allow CRLF in front of first boundary even in strict mode

  Browsers do not do this, but some HTTP client libaries do and it's
  technically allowed.

fix: Fail on invalid first boundary instead of skipping the first segment

  As per spec, the start boundary must be at position zero, or start with CRLF
  to separate it from the preamble. Boundaries are forbidden in segment
  bodies, but not in the preamble. This means that a preamble can actually
  contain the boundary, as long as it does not start with CRLF. This is
  nonsense, so let's ignore the spec here. A preamble that contains the
  boundary is so rare and suspicious that we assume a broken or malicious
  client and fail fast, even in non-strict mode. This is way better than
  silently skipping the first segment and loosing data.

fix: Accept tiny (1 byte) chunks and arbitrary chunk borders

  There were edge cases where a very uncommon chunk border would break the
  parser. This is fixed and properly tested now.

fix: Fail if message did not contain the boundary at all.

  In strict mode, we fail very fast (because the message is either empty, or
  the message does not start with a boundary. In non-strict mode, we fail
  at the end when the parser closes in PREAMBLE state.

We deliberately break spec and reject input that is technically valid. The test that ensures this should be clear on why that's a good idea.