feat: add absorb for in-place merges of buffered digests

[kei-nan]

Folding many small digests into one, for example per-batch digests built with push, currently means flushing each one and calling merge_digests. That sorts, compresses and allocates for every digest, even when it holds only a few dozen values.

This PR adds TDigest::absorb(&mut self, other: TDigest):

• If other has never flushed, its buffered values are pushed into self. The result is exactly what pushing them into self directly would give.
• Otherwise, both digests are flushed and combined with merge_digests.
• The result keeps the larger max_size. As with push, call flush before queries.

This is a new public method, so it's a minor-version addition with no existing behavior changed.

Tests

• Unit tests: absorbing buffered chunks equals pushing directly (full PartialEq); absorbing a flushed digest equals merge_digests; empty digests in both positions; the larger max_size is kept.
• Proptest: absorbing buffered chunks of any size below the flush threshold equals pushing directly.
• Forcing absorb onto the merge path makes the buffered-path tests fail.

Bench and docs

• New absorb_small_digests bench: 10,000 values in 64-value digests, absorb vs. flush + merge_digests.
• README.md and docs/architecture.md describe the method, and the cost table has a row for it.

Checked locally: cargo test (debug and release, all feature sets), clippy -D warnings, rustdoc -D warnings, no_std, and the MSRV 1.62 checks from CI.

This is independent of the integer-key sort PR. If both land, absorb pushes through whichever buffer representation is in place.

🤖 Generated with Claude Code