fix: add sleep to fibonacci test

I hadn't added a sleep to the fibonacci test previously because the delays
added by the math being performed has always been sufficient to ensure that
lines sent to stdout and stderr don't arrive out of order. But apparently
the hydra.nixos.org build farm has faster machines because one managed to
trip the race condition: https://hydra.nixos.org/build/290187717/log

Adding 10ms sleep to make it signficantly more likely that lines sent to
stdout and stderr will arrive in the right order.

chore: update Makefile and flox environment

Update flox environment, fix version handling and tidy Makefile in
preparation for adding to nixpkgs upstream.

fix: move midline-flush.c sleep to occur between flush() invocations

We were doing the sleep() invocation between writing to the stdout/err
file descriptors rather than between the times we were flushing them.
This update divides the time slept by 3 (bringing it to approx 3x what
it was before we encountered CI failures), while moving the sleep between
the two flush() invocations.

It also adds an extra sleep following the second flush() invocation to
make sure that the kernel has time to communicate its contents before
the subsequent test.

fix: increase sleep in midline-flush.c (#11)

* fix: increase sleep in midline-flush.c

Increasing sleep tenfold to really minimize the chances of test races at build time.

fix: expect less of kernel, sleep longer in tests

Observed repeated test failures on CI running on aarch64-linux instances:

  > Failed test tests/100-lines
  > --- tests/100-lines.log      1970-01-01 00:00:01.000000000 +0000
  > +++ /build/tmp.Xnq7ZeMG7J/100-lines.log    2025-01-29 15:07:16.278776269 +0000
  > @@ -49,8 +49,8 @@
  >  line 25 to stdout
  >  line 25 to stderr
  >  line 26 to stdout
  > -line 26 to stderr
  >  line 27 to stdout
  > +line 26 to stderr
  >  line 27 to stderr
  >  line 28 to stdout
  >  line 28 to stderr
  > make: *** [Makefile:137: tests/100-lines/log] Error 1

This occurs when the kernel _actually_ delivers messages out of order to
the stdout and stderr file descriptors, which is very much possible for
the Linux kernel.

There's really no resolution for this apart from lowering our expectations,
and this patch roughly triples the time spent sleeping between lines sent to
stdout/stderr by changing `sleep 0` to `sleep 0.01`:

  $ time sleep 0
  real    0m0.003s
  user    0m0.004s
  sys     0m0.000s

  $ time sleep 0.01
  real    0m0.013s
  user    0m0.002s
  sys     0m0.001s

fix: handle polls with both POLLIN and POLLHUP (#9)

The behaviour of poll() seems to differ between Linux and Darwin when the
child process has closed the pipe and poll() returns both POLLIN and POLLHUP
set on the returned events value. On Linux this is a sign that we can read
from the pipe one more time, but on Darwin it seems that POLLIN remains set
after the final read so we need to add extra logic to make sure to stop
polling for POLLIN after a failed read in the presence of POLLHUP.

Also fix minor formatting issues introduced by Devin.

fix: address flushes that happen in the middle of a line

Simplify the logic around capturing output from exec'd process. Swap
the use of poll() with the more straightforward read() idiom in the
(relatively simpler) stdout & stderr worker processes and add support
for relaying lines received across multiple read() system calls. Also
add test for testing lines received across multiple read calls.

Also updated ANSI color palette and default colors.

feat: support relative timestamps

The timestamps in build logs are often more useful when measured from the
beginning of the t3 invocation rather than displaying the wall clock time.

fix: return RC of invoked command, add tests

Initial commit