Phase 7: conditional formatting

[billdenney]

Part of the libxlsxwriter API coverage roadmap in #112.

Covers worksheet_conditional_format_cell() and _range() through four
constructors, since the struct's ~30 fields fall into clusters that share almost
nothing:

xl_sheet(df, conditional = list(
  xl_cond_cell("A2:A100", criteria = ">", value = 100,
               format = xl_fill(background = "#FFC7CE")),
  xl_cond_scale("B2:B100", colors = c("red", "yellow", "green")),
  xl_cond_bar("C2:C100", color = "steelblue", solid = TRUE),
  xl_cond_icons("D2:D100", style = "3_traffic_lights")
))

Two silent-wrong-output traps closed

The 34 criteria are partitioned by rule type, and libxlsxwriter checks none of
the pairing. A text criteria on a cell rule produces a file Excel accepts and
then silently ignores. .CONDITIONAL_CRITERIA_FOR maps each type to what it
accepts; a mismatch errors naming both and listing what would be valid there.
Where the two are redundant, the type is inferred from the criteria.

Three of libxlsxwriter's type constants carry no TYPE_ in their names —
LXW_CONDITIONAL_2_COLOR_SCALE, _3_COLOR_SCALE, _DATA_BAR. My first reading
of the header missed them and put icon_sets at 14 instead of 17. Fourteen is
the index of a two-colour scale, so every icon set would have silently written
a colour scale. Caught by going back to the header rather than trusting the
earlier extraction.

The fix is pinned by enumerating all seventeen styles against the names Excel
stores — 5_quarters → 5Quarters, 4_ratings → 4Rating — which is exactly
the test that would have caught the original error. A further test asserts the
map still has seventeen entries, so a style added by a libxlsxwriter update
fails loudly rather than becoming quietly unreachable.

The consistency gate earned its keep

xl_cond_cell() was written with criteria before type, while
xl_validation() has type before criteria — six shared arguments in
conflicting order. The gate added in #127 failed the build before this shipped.
Reordered to match, on the principle that in both functions the type is what
determines which criteria are legal.

Verified, not assumed

• <dxfs>: conditional formats use differential formats rather than
  cellXfs, so I could not assume the shared format registry would land
  correctly. It does — a rule's fill appears in <dxfs> and the rule references
  it by dxfId. Pinned by a test.
• Icon sets embed nothing — Excel draws them itself. A test asserts no
  xl/media part appears, which is why they did not need to wait for Phase 10.
• No constant-memory flip — no optimize guard in either function; a test
  asserts row streaming stays on.

Verification

• 1241 tests pass, 0 failures, 1 pre-existing skip
• devtools::check(): 0 errors, 0 warnings, 0 notes
• Assertions against observed XML for all four clusters: cellIs with each
  operator, colorScale with 2 and 3 stops, dataBar, iconSet
• Boundary cases: criteria illegal for its type; criteria supplied where the
  rule carries its own meaning; between missing a bound; scales with 1 or 4
  colours; mismatched values/rule_types lengths; all 17 icon styles