refactor(output): measure table cells once and fit columns in one flat pass

[ryanlewis]

This is step two of the internal/output consolidation, following #314. It changes how table is put together and keeps every byte of output the same.

What fit() does now: it measures every column once, then runs one step after another, each stopping as soon as the row fits:

1. Drop the dropFirst columns.
2. Cut the shrink columns to their soft floor.
3. Switch the dropOrder columns to their compact forms.
4. Drop the dropOrder columns.
5. Give back any room steps 3 and 4 freed beyond what was needed (restore).
6. Cut the shrink columns to their hard floor.

Before, the same order was spread across compactTier, drop, expand, regrow and two shrinkTo calls. Those passed three parallel []bool slices and a widths slice that each step changed, and needed freed flags, a guard on running expand and regrow, and bounds checks in four loops.

Changes:

• A row is a slice of cell{text, tail, alt}, measured once when the row is added. The tail()/alt() setters, and the [2]int-keyed maps behind them, are gone. printTasks builds its tag, title and date cells and adds the row in one call.
• Each column has one state struct, col, in place of the parallel slices.
• drop and compact are two small steps that untilFits runs in turn, which leaves fit() at 13 lines (it was 26).
• expand and regrow are merged into restore, which always runs. The old guard on whether anything had been compacted or dropped never changed the result.
• The soft and hard shrink passes share shrinkTo and differ only in their floor function.
• measure creates state for every column the drop and shrink lists name, so fit needs no bounds checks. A negative index is a bug and panics; before, it was ignored, and no printer uses one.
• padCol, joinWithGap and dropEmptySpans are unchanged.

table.go goes from 429 to 388 lines, its non-comment code from 293 to 256, and its functions from 22 to 20.

How I checked the output is identical:

• render_golden.txt is unchanged. Since test(output): pin piped, wide-rune, dim-cut and tab/newline title rendering #314 it covers piped output too.
• The old table is kept as legacyTable in table_legacy_test.go. TestTableMatchesLegacy (seeded) and FuzzTableMatchesLegacy render random tables through both and compare each line byte for byte at every width up to 130. The tables use the printers' own column setups and random ones. Cells cover dim text styled one character at a time, wide characters, tabs, newlines, tails, compact forms and short rows. I fuzzed for 10 minutes (272k runs), then another 3 minutes after the review changes, with no difference. Reversing the order restore regrows columns in makes the test fail at once. The next PR removes both files.
• I built things from main and from this branch and ran both against a Things backup database, with HOME pointed at a temp directory. The commands were today, inbox, upcoming, anytime, someday, repeating, logbook, trash, deadlines, projects (open and completed), areas, tags, two searches and two shows. Each ran piped and under a script pseudo-terminal at 40, 62, 80, 100, 120 and 200 columns, with --color never and --color always. All 238 pairs were byte-identical. script sometimes drops a \r, which happens even comparing main against itself, so pseudo-terminal output is compared with \r removed; piped output is compared raw.
• make test and make lint pass. While the differential test is in place, the output package's tests take about 19s under -race.