Add createBloatedToastTable helper with realistic ~10KB/row data
(5 chunks per value, STORAGE EXTERNAL) and 7 test cases covering
basic/high/low/no bloat, small tables, CLI output, and chunk sampling.

- Sample full-size chunks (chunk_seq > 0) with fallback to avoid
  partial last-chunk bias in ppc_ref calculation
- Filter out tables with relpages = 0 (stale/uninitialized stats)
- Use current_setting('block_size') instead of hardcoded 8192
- Move helper function to pg_temp schema for crash-safe cleanup
- Add stale stats warning when no VACUUM in the last 24 hours
  (joins pg_stat_user_tables for last_vacuum/last_autovacuum)
- Align sql/toast_bloat.sql reference file with embedded query

TOAST tables do use the FSM for new insertions (via heap_insert),
contrary to what was previously stated. The reason UPDATE col=col
doesn't compact TOAST is that PostgreSQL reuses the existing TOAST
pointer when the value hasn't changed, not because TOAST is append-only.

The previous chunk_seq > 0 sampling strategy returned the partial last
chunk for values with exactly 2 chunks (e.g. ~3KB data), producing a
~48% false-positive bloat estimate on non-bloated tables.

New strategy: find a multi-chunk chunk_id first, then read its
chunk_seq=0 which is always exactly TOAST_MAX_CHUNK_SIZE (1996 bytes
on 8kB blocks). Falls back to any chunk for single-chunk-only tables.

Adds TestToastEstimateTwoChunkValues regression test.

More readable output for standalone usage: min_pages_required and
bloat_pages are directly comparable to toast_pages, consistent with
the min_pages_required pattern in sql/table_bloat.sql.

The embedded Go query is unchanged (uses raw values internally).

Compute TOAST_MAX_CHUNK_SIZE as block_size/4 - 52 instead of
sampling actual chunks via PL/pgSQL function. Uses constants CTE
pattern matching table_bloat.sql style.

Add --toast and --heap flags to command reference, quick start
examples, output sample, and bloat estimation explanation.

Replace hand-crafted README output examples with real qwash output
from a demo database. Fixes order_items at 41% shown in CRITICAL
(≥50%) section. Examples now show all severity levels (CRITICAL,
HIGH, MEDIUM) with correct categorization.

Add omitempty to BloatReport.Indexes JSON tag to suppress
"indexes": null from JSON output when no index data is present.

Add toast field to BloatReport JSON structure so --toast --json
produces proper TOAST bloat data instead of empty output. Works
for all combinations: --toast alone, --heap --toast, and with -t
table filtering. Add TOAST JSON example to README.

The sampling function was reading only chunk_seq=0 (always full-size
at TOAST_MAX_CHUNK_SIZE), ignoring that the last chunk of each value
is typically much smaller. This caused systematic underestimation.

Fix: average all chunks of a sampled value to capture the real
per-chunk size. On test data, TOAST bloat estimation goes from
55% to 69.6%, matching pgstattuple reference (69.5%).

Replace single-value sampling with avg(length(chunk_data)) across
all live chunks. This naturally handles multi-column tables and
mixed payload sizes with ~90 ms per 100 MB of TOAST data.

Replace octet_length() with pg_column_size() which reads va_extsize
from the TOAST pointer. Unlike octet_length() which returns the
decompressed size, pg_column_size() returns the actual stored size,
making it correct for both EXTERNAL and EXTENDED storage.

Rewrite standalone SQL as DO block + named cursor with dynamic
UNION ALL of pg_column_size() per table. No objects created.

Add toast_compressed table to demo for EXTENDED storage testing.

Use exact multiples of TOAST_MAX_CHUNK_SIZE (1996 bytes) in test
payloads to avoid partial last chunks that skew avg(length(chunk_data))
and cause false bloat overestimation with the full avg scan approach.

- CHANGELOG: Add v0.3.0 section with TOAST estimation feature
- Project docs: Add toast_bloat.go, toast_bloat.sql, --toast/--heap flags,
  update implementation status and TODO list
- sql/embed.go: Clarify that toast_bloat.sql is standalone, not embedded
- README: Precise TOAST estimation method (pg_column_size, not constant)
- goreleaser: Include standalone SQL queries in deb/rpm packages

Already in .gitignore, should not be tracked.

- Bump version references from 0.2.0 to 0.3.0
- Fix package filenames: add linux_ prefix to match goreleaser output
- Update TOAST output examples with realistic data