Decide row streaming per workbook, and say when it costs something

[billdenney]

Adds constant_memory and constant_memory_threshold to write_xlsx(), and
replaces "stream unless a known-bad feature is present" with an explicit,
ordered decision.

They sit on write_xlsx() rather than xl_properties() because streaming
governs how the file is produced, not what the document contains.
use_zip64 is the same kind of knob.

The rules

1. constant_memory = FALSE → off, without evaluating anything else.
2. A blacklisted feature → off, absolutely. TRUE cannot override it, and
   is warned about instead.
3. constant_memory = TRUE → on, overriding the size estimate.
4. Estimated extra memory below constant_memory_threshold (default 128 MiB)
   → off.
5. Otherwise → on.

Rule 2 is deliberately not overridable. Three of the four blacklisted features
fail loudly under streaming — tables are refused outright, merges and embedded
images hit dimension checks. But a multi-cell array formula writes a file
that opens cleanly in Excel and is quietly missing cells, which is
jmcnamara/libxlsxwriter#522, closed upstream as won't-fix. No warning makes
that acceptable, so the request is refused rather than honoured.

Rule 4 is new. Streaming only pays for itself on large data, and not streaming
also produces a smaller file, since repeated strings go to the shared string
table instead of being written inline.

Where the estimate comes from

110 bytes per cell, measured rather than guessed — against nycflights13 at
three sizes and three column mixes:

shape	cells	bytes/cell
repetitive text	200k → 2M	37 → 82
numeric	700k → 7M	79 → 107
mixed	950k → 9.5M	87 → 112

Repetitive text is cheapest, because the shared string table deduplicates it.
The high end is used, so the estimate errs towards keeping streaming on — which
at scale is the faster path as well as the leaner one.

Cells are summed across the workbook, not per sheet: libxlsxwriter holds
every sheet's cell table until the file is closed.

What streaming actually costs

Peak process memory, staged so R's share can be separated from the C library's.
The prep row runs everything write_xlsx() does up to but not including the
.Call:

stage	337k rows	1.01M rows
R + package, no data	92 MB	93 MB
data frame loaded	253 MB	365 MB
all R-side resolution, before .Call	253 MB	463 MB
write, streaming on	253 MB	458 MB
write, streaming off	931 MB	2,608 MB

Essentially all of the difference is libxlsxwriter's cell table: R's whole
contribution is 253/463 MB, streaming adds nothing measurable, and not
streaming adds 678 MB and 2,145 MB. At 1M rows the C library wants 4.6× what
R does, and takes 65s against 27s.

(Measuring this needs bin/x64/Rterm.exe — both Rscript.exe and bin/R.exe
are launchers and report a 7 MB shim.)

The suggestion

When a blacklisted feature turns streaming off and the workbook is large
enough for it to matter, write_xlsx() now says so, naming the feature and the
estimated cost. Silent below the threshold, where there is nothing to act on.

Test churn, and why

The threshold changes the default for small workbooks from streaming to not,
which moves strings from inline to the shared table. Ten tests asserted the
inline spelling.

Each now passes constant_memory = TRUE, because that is what they were
actually testing — does this feature force streaming off? — a question the
size rule had been answering for them by accident. request = TRUE isolates
the feature's effect from the size rule, which is a sharper test than before.

Also included

Two things that rode along with this work rather than belonging to it:

• Sheet-qualified ranges. .xl_resolve_range() now parses Data!A1:B5,
  'My Sheet'!A1:B5 and 'It''s'!A1, and a sheet element in the list form.
  It is opt-in via allow_sheet: an autofilter, merge, validation,
  conditional format, table, image anchor and print area all apply to the sheet
  they are on, so a sheet name there is a misunderstanding to report, not a
  refinement to honour. Only chart series can genuinely point elsewhere, which
  is what this is for. Everything else now rejects one with a clear message
  instead of ignoring it.

• lxw_table_column.format is ignored for data written by the caller jmcnamara/libxlsxwriter#520 resolved. Not a bug — a documented
  limitation, and the maintainer's recommended fix is what writexl already
  does. The actionable detail is that both the column plan and the table
  column's own format are needed, the latter becoming dataDxfId, "for strict
  correctness with Excel". A test now pins both, because after the column-plan
  fix the dataDxfId looks like the redundant one.

Verification

• 1747 tests passing, 0 failures, 0 warnings.
• R CMD check: 0 errors, 0 warnings, 1 NOTE — the pre-existing spelling
  wordlist.
• 13 files changed, +275 / −27.