Before: `DEVOURER_DUMP_CANARY` emitted one fixed register list for
every Jaguar chip and `canary_kernel_dump.sh` mirrored that list.
Two problems:
- 8821AU is 1T1R: path-B BB-AGC mirror (`0xe20-0xe40`) and RF[B]
  reads come back as BB-init defaults or sentinel zeros. Including
  them in the dump cluttered the diff with ~16 always-divergent
  noise rows (16 lines on 8821, of which 12 were path-B BB regs
  and 6 were RF[B] sentinels).
- 8814AU additions (PR #72) bolted path-C/D BB regs onto the
  end as a conditional block instead of integrating them into a
  proper chip-aware schema.

This PR refactors into per-path register groups (`bb_anchors`,
`bb_pathA`, `bb_pathB`, `bb_pathC`, `bb_pathD`, `mac_canary`,
`rf_canary`) with chip-dependent path masks:

- `CHIP_8821`: path-A only — both BB path-B mirror and RF[B] skipped.
- `CHIP_8812`: paths A + B (unchanged from prior behaviour).
- `CHIP_8814A`: paths A + B + C + D BB-table; RF still A + B only
  (paths C/D are HW write-only).

Per-chip line counts on a clean fresh capture at ch6:
- 8821AU:  52 lines (was 68; -16 sentinel/no-op rows)
- 8812AU:  68 lines (unchanged)
- 8814AU:  90+ lines (full path-A/B/C/D coverage)

`tools/canary_kernel_dump.sh` accepts `8812` / `8821` / `8814` for
the chip arg and emits the matching path subset so kernel + devourer
dumps still diff line-by-line.

No functional behaviour change — purely diagnostic output shape.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>