Make LinearRecurrenceFormula truly O(1) memory in NaturalArray formula path; bump to v0.8.4

[Copilot]

Linear recurrence formulas were still carrying range-sized precomputation overhead, unlike the now O(1) conditional formula path. This PR removes recurrence precompute tables and aligns recurrence formula composition with the same constant-space, newer-over-older model.

• Linear recurrence memory model

  • Removed eager range-table precomputation from LinearRecurrenceFormula (buildPrecomputedValues path).
  • Kept direct evaluation via recurrence math per lookup, avoiding storage proportional to covered index span.

• Recurrence composition parity with conditional formulas

  • Added LinearRecurrenceFormula.compose(newer, older).
  • Added composite-aware contains(...) and evaluate(...) so recurrence formulas resolve in newer-first order over overlapping ranges.
  • Updated NaturalArray.addLinearRecurrenceFormula(...) to merge into a single composed recurrence formula instead of appending unbounded entries.

• NaturalArray recurrence cache cleanup

  • Removed per-index recurrence-formula index cache usage, eliminating linear growth in metadata from random recurrence reads.

• Release/version updates

  • Bumped repository version to 0.8.4.
  • Added v0.8.4 changelog entry focused on the recurrence O(1) improvement.
  • Updated visible version strings to 0.8.4 in README/docs/install fallback.

Example (new composition model):

if (linearRecurrenceFormulas.isEmpty()) {
    linearRecurrenceFormulas.add(formula);
} else {
    int lastIndex = linearRecurrenceFormulas.size() - 1;
    LinearRecurrenceFormula current = linearRecurrenceFormulas.get(lastIndex);
    LinearRecurrenceFormula merged = LinearRecurrenceFormula.compose(formula, current);
    linearRecurrenceFormulas.set(lastIndex, merged);
}