[FEATURE] Shrink the on-disk size of the DiskSeismic indexes

[chishui]

Problem

A disk_seismic(_sq) .dat runs several times the size of its corpus, much of it carrying no
information. The inline forward index dominates; duplicating each document's vector per block is
the design, not the issue. On top of that:

• Every block is padded to a page (kPageAligned, kDefaultPageSize = 4096 —
  io/inline_forward_index_io.h:82,92) yet holds roughly ten documents at standard
  lambda/beta, and a query never reads the padding. InlineLayout::kPacked already exists
  (inline_forward_index.h:46); the header records the alignment.
• [u16 comps[total_nnz]] (:29) is flat, though a document's ids ascend and so compress.
• doc_id[], off[] (:113-114) are uint32_t, beyond what a block needs.

Proposal

1. Default to kPacked: no new encoding, no format bump.
2. Add byte accounting, so later steps are sized rather than guessed.
3. Then evaluate separately, each kept only if latency holds: delta-coded ids, narrower off[],
   4-bit values.

Acceptance: smaller .dat, recall and latency unchanged.

[chishui]

resolved by #60