feat(bfmatcher): brute-force Hamming matcher and match_t

Summary

Add src/bfmatcher/bfmatcher.ts — a brute-force Hamming matcher for binary
descriptors — plus the match_t correspondence type.

This is the first half of #83, extracted so it can land independently.
#83 narrows to pose_estimator.

Why

jsfeatNext can detect keypoints and compute ORB descriptors, but it cannot match
them. The only matcher in the repository is match_pattern(), written inline in
examples/sample_orb_pinball.html. A consumer of the published package
therefore cannot run the natural-feature pipeline end to end without
reimplementing matching themselves.

Two further things depend on this:

• GMS (separate issue) is a post-matching filter. Without a matcher and a
  match_t type it has nothing to operate on.
• TEBLID (separate issue) produces 64-byte descriptors at p512. The matcher
  must therefore handle arbitrary descriptor widths from the start, not the
  32 bytes the sample hardcodes.

Extracting this from #83 keeps pose_estimator — the larger, more speculative
half — off the critical path.

Scope

#83 already carries a full prototype in its body; this issue is its
implementation, with one correction (generalised width).

• match_t { queryIdx, trainIdx, distance }, the equivalent of cv::DMatch.
• match(query, train, maxDistance) — nearest neighbour, with optional
  crossCheck.
• knnMatch(query, train, k) — k nearest neighbours per query descriptor.
• ratio_test(knn, ratio) — Lowe's ratio test helper.
• Hamming distance via SWAR popcount over the Int32Array view, bit-identical
  to the sample's popcnt32.
• Generalised descriptor width (columns a multiple of 4), not hardcoded to
  32 bytes.
• Attached to the namespace as jsfeatNext.bfmatcher (singleton, per the
  post-0.9.0 convention) and jsfeatNext.match_t (constructor).
• NORM_HAMMING mirrors OpenCV's constant value.

Acceptance criteria

• match, knnMatch, ratio_test and crossCheck implemented, with the
  OpenCV-shaped API of the #83 prototype.
• Numeric parity with the inline match_pattern() of
  sample_orb_pinball.html: same input descriptors produce the same
  correspondences and the same distances.
• Works at 32 and 64 bytes per descriptor, so TEBLID p512 is supported
  the day it lands.
• Buffers borrowed from the shared pool are balanced (get_buffer /
  put_buffer), covered by the pool-balance test helper.
• No new any (see #85).
• Attached to the namespace and covered by the API-shape test.

Out of scope

• pose_estimator and homography-to-pose decomposition — stays in #83.
• FLANN or any approximate nearest-neighbour index.
• Match filtering / geometric verification — see #134.
• Float descriptors and L2 distance: NORM_HAMMING is the only norm here.

Related

• Split from: #83 (which narrows to pose_estimator)
• Consumers: #134 (GMS), and #135 (TEBLID) for end-to-end matching tests
• Contract: #128 (Match type and match kind guard), #129 (filterMatches)
• Plan: docs/features2d-expansion-plan.md