shuf: match GNU's RNG and input handling

[metif12]

Fixes #189 and #193.

This branch contains only the shuf change, so the diff is limited to src/shuf/.
It is the third of a series and should be merged after #195 and #196:

• build: make all 76 utilities compile against current V #195 gets the build and make testfmt green again
• fix: repair the 10 test files that were failing for unrelated reasons #196 repairs the other 10 failing test files
• this one finishes shuf

With all three applied the branch is green end to end: v run build.vsh builds
all 76 utilities, make testfmt passes and v test . is 47/47, up from
37/47 on main.

--random-source produced the wrong permutation (#189)

The old code read the random source into memory, summed the bytes, fed that to
a V PRNG, and shuffled with a hand written swap loop. That is not what GNU does,
so the same file gave a different permutation, and the test that checks it
failed.

src/shuf/randint.v is now a direct port of GNU's gl/lib/randint.c and
gl/lib/randperm.c. The leftover entropy in randnum/randmax has to be
carried across calls exactly as GNU does, because that is observable:
shuf -i 1-5 over a one byte source still consumes a single byte, and only
because of that carry.

GNU also switches to a hash based sparse representation for large sparse
permutations. That is purely a memory optimisation — it performs the same
swaps in the same order — so it is not reproduced here.

Verified byte for byte against GNU coreutils 9.4 for:

-i 1-5                -i 0-99 (exhausted)     -n 3 from 5 lines
all lines             -r -n 6 (echo/file/range)  -o
stdin                 -n larger than input    missing newline / missing NUL
empty lines kept      -z

and for the end of file, No such file or directory, invalid input range,
extra operand and cannot combine -e and -i errors.

Other behaviour that was wrong

• Empty lines were dropped. register_lines_by_file had
  if line.len > 0, so shuf silently lost them.
• -z did not work. The zero-terminated path appended the whole input as a
  single element instead of splitting on NUL, and open_operands split on
  newline regardless of the flag.
• -n 0 output every line instead of nothing. The old code used 0 to mean
  "every line", which is exactly what -n 0 is not.
• -r -n COUNT re-shuffled for every line of output, so it could repeat
  the entire input COUNT times instead of picking COUNT lines.
• -i accepted anything. split('-') with no validation, so 5-1,
  abc, 1-x and 1-2-3 were all accepted.
• -e together with -i, and extra operands, were accepted.
• A missing input file reported V's failed to open file "..." instead of
  GNU's No such file or directory.

Help text (#193)

The option descriptions now use GNU's wording, and the usage lines match GNU's:

Usage: shuf [OPTION]... [FILE]
   or: shuf -e [OPTION]... [ARG]...
   or: shuf -i LO-HI [OPTION]...

The options use the _opt flag variants so V stops appending a
(default false) note to each one, and the value names are LO-HI, COUNT
and FILE as in GNU:

  -e, --echo                treat each ARG as an input line
  -i, --input-range LO-HI   treat each number LO through HI as an input line
  -n, --head-count COUNT    output at most COUNT lines
  -o, --output FILE         write result to FILE instead of standard output
  --random-source FILE      get random bytes from FILE
  -r, --repeat              output lines can be repeated
  -z, --zero-terminated     line delimiter is NUL, not newline

The surrounding boilerplate is still whatever common.flag_parser produces,
because that is what every utility in this repository uses and the README asks
for consistency. Reproducing GNU's exact --help layout would mean bypassing it
and diverging from the other 75 tools. The descriptions, which is what the
issue was about, now match.

Tests

test_random_source no longer hardcodes a permutation. The expected value
1 4 5 2 3 came from GNU 8.32 and does not match any current GNU — 9.4 gives
5 2 1 3 4 for the same input, and I checked that the port reproduces 9.4
exactly. Since the permutation for a given random source is fully determined,
the tests compare against the platform shuf through the rig, which is what
the other tests in this file already do and keeps them valid across GNU
versions.

Added coverage for lines, -n, repeat, exhaustion, a missing source, empty
lines, -z, -o, -n 0, and the input-range and operand errors.

One of the new tests caught a bug while I was writing it: open_operands split
on newline unconditionally, so -z was still wrong after the first round of
fixes. Fixed by threading the delimiter through.