sort: implement GNU's orderings, keys, check mode and file handling

[metif12]

Closes #192.

sort sorted each input file on its own and concatenated the results, offered five orderings of the eleven GNU has, had no working --stable, and reported a missing file as fn (string) IError. This implements what GNU's sort 9.4 does.

One stream, not one sort per file

sort f1 f2 sorts the inputs together. The old code sorted each file separately and appended them, so sort f2 f1 could report an order the files did not have. -m is the option that means "each input is sorted already", and it now merges the streams without sorting them, so a line out of place in its own file stays there.

Reading also changed: stdin no longer goes through a temporary file, -z makes the NUL the record separator, and --files0-from reads the list of names. The line number a disorder is reported at is counted while reading rather than by comparing the sorted output with the input.

The orderings

-g, -h, -M, -V and --sort=WORD were missing, and the one -n had was GNU's general numeric one. Each is now a separate grammar, because GNU's are not interchangeable:

• -n takes blanks, a minus and digits. A leading + weighs nothing, so +5 sorts as zero.
• -g takes what strtod does, including a leading +, an exponent and a hexadecimal float, and weighs a line with no number as minus infinity. The words inf and nan are recognised, since V's atof64 returns 0 for both.
• -h adds a size suffix and takes an exponent only with a capital E. Its units are capital-only from P upwards: GNU weighs 1p as one and 1P as a petabyte.
• -M weighs a month name and then the text after it, so JAN precedes JANUARY.
• -V weighs digit runs as numbers. Its character order is not the byte order: sorting every printable character as a one-character name shows a dot and a tilde first, then digits, then letters, then the rest — but only at the start of the key, since ab sorts before a.b.

After equal numbers or equal months, GNU does not weigh the rest of the line; the last-resort comparison of the whole lines does that. -s turns the last resort off, and so does -u, which is why -u -n keeps the first line of a run of equal numbers in the order the lines arrived in. V's sort is not stable, so the position of each line is part of the comparison.

Keys

-k took its ordering letters from only one place and read a field as if -t were always in effect. A key is now read as F[.C][OPTS][,F[.C][OPTS]] with the letters allowed on either side of the comma, so -k2n,2 and -k2,2n agree.

Without -t, a field starts at the whitespace before it, and the first field has none to start from — which is what makes a line of nothing but blanks have an empty second field. -b steps past that whitespace. A character position is the last character of the key counted from the start of its field, and a line shorter than that stops early. A field number of zero is rejected, as at GNU, rather than read as the end of the line.

Checking and messages

-c reports FILE:LINE: disorder: LINE and stops at the first one; -C says nothing. With -u a line whose key equals the one before it counts as a disorder, and -c takes only one input. The messages for a bad key, a multi-character -t and an extra operand with -c now match GNU's, including which of them get the advice about --help and which quote the name.

Testing

A table of 43 cases and a randomised matrix of 413 invocations over six sets of awkward inputs were compared against GNU, across every ordering, key form, check mode and error path. Five seeds give 413 of 413 on three of them and 409 of 413 on the other two.

The four that differ are all -h weighing a plain number of seven digits or more against a suffixed one: GNU puts 1000000 before 1K, which no value for the K is consistent with, since it would have to be above 1e7 and at most 1024^4. The reasoning is in number.v. Every other comparison of a suffix against a plain number agrees, including 1K against 1024 and 2K against 1M.

The tests were rewritten as well: they tested sort(options) and parse_sort_key, both of which are gone, and three of them asserted an order GNU does not produce.

Stacked on #195, #196, #197 and #198, v fmt -verify . passes, v run build.vsh builds all 76 utilities, and v test . reports 47 of 47 files passing.

Note for review: this branch is based on main, so its own CI run will still show the pre-existing failures that #195 and #196 fix, and the v fmt gate will report the files #195 reformats.