perf: cut per-node and per-die work in the evaluator #336

[edloidas]

Collapsed EvalContext's expression/rendered fragment arrays into single strings, then moved both onto EvalResult. Every eval* runs against a context of its own and pushed each array exactly once, so the arrays never held more than one element and every parent joined a one-element array. Returning them instead makes a forgotten write a type error rather than an empty string in a user-visible RollResult.expression — it caught a missing evalFunctionCall field on the first compile.

EvalContext now holds only what flows up unchanged or merged (rolls, versusMetadata) plus the parent-set suppressRender; everything a wrapper reformats rides EvalResult.

Rebuilt rewriteFlags as one loop instead of a closure, a filter array and a spread. It still always returns a fresh array — a variant returning the input when it already equalled the result was rejected because mergeMetaRolls clones a die as { ...die, modifiers: rewriteFlags(...) } and would alias flags that later passes mutate.

Suppressed the discarded pool render in a modifier's target context. Each pool modifier rebuilds the breakdown from its target's expression plus a renderDice pass of its own, so the target's own render was thrown away; in a chain each modifier's context is the next one out's discarded target, so every level above the innermost was also rendering into the void.

Also swapped keptIndices.includes for a Set in renderGroup, added 100d10ro<3 and 100d10kh50sd cs>8 benchmark cases for the two shapes with no coverage, and recorded why toSeedString stays on decimal.

Measured, medians of 5, A/B against master:

notation	before	after
1d20	0.353 us	0.277 us	-22%
2d6+3	0.692 us	0.530 us	-23%
4d6kh3	1.256 us	1.034 us	-18%
floor((1d4+1)*2/3)	1.823 us	1.273 us	-30%
2d20kh1 vs 15	1.440 us	1.098 us	-24%
10d10sd>=6f1	3.185 us	2.612 us	-18%
100d10kh50sd cs>8	47.393 us	39.184 us	-17%
60-term sum	21.192 us	14.080 us	-34%

No behavior change: rendered, expression and the RollResult shape are byte-identical, and the renderBreakdown parity property test still holds. Sizes fell rather than rose, so no budget change: index.js 13.16 -> 13.09 kB, { roll } 12.91 -> 12.85 kB.

Closes #336