Fixes #40005.

Fixes: #40004

This test only ensures that the new fields do
not cause the parser to crash, it does not look
at actually parsing those fields.

Fixes: #40003

Implements #40002.

When parsing a large file with many descriptors we would repeatedly
search the remaining file for the sequence "newline + keyword + space"
and then "newline + keyword + newline" to find the start of the next
descriptor. However, if the keyword is always followed by newline, the
first search would always fail.

The optimization here is to search once whether the keyword is
followed by space or newline and avoid unnecessary searches when going
through the file.

In the long term we should use a better parser. But in the short term
this optimization will have a major impact on performance, in
particular with regard to concatenated microdescriptors.

Implements the library part of tpo/metrics/statistics#40002.

Implements tpo/metrics/library#40001.

Implements the remaining part of #30602.

Implements #26673.

As discussed on #34031, we should avoid keeping a map of known failure
types in the sources and simply put together the error code in the
format we want.

This patch implements that change.

We are detecting descriptor types of parsed descriptors by either
content or file name. In some cases, if we downloaded descriptors from
web servers, there is no file name. In other cases the file name can
match more than one descriptor type. It seems most robust to move the
file name checks to the end, which includes web server access logs and
OnionPerf analysis files.

It seemed like a good idea to decompress xz-compressed files before
handing over decompressed bytes to descriptor implementations.
However, our web logs expect descriptor bytes to be in compressed
form, and now they're unhappy about receiving uncompressed bytes.
Changing this back and teaching OnionPerf analysis files to expect
compressed bytes, too. Found by running CollecTor's unit tests.

Implements #34070.