[AIMOD-1354] Add "delete" index for AMP keywords

[vazish]

• This PR adds a "delete" index for the "full_keywords" field from AMP json.
• This will be used as part of the treatment branch for an A/B AMP experiment in Merino with the fuzzy matching portion as a treatment branch.
• The best place to start review is SymIndex.build() function that builds the delete index over the full keywords, i.e. generates for the keyword example "world cup" -> "orld cup", "wrld cup", "word cup, ...
• Then query_fuzzy() for the how it's called.
• Once merged, I might need some help figuring out how to release!
• (I did use Claude's help with the Rust side but reviewed it to make sure the approach was what I had in mind)

Some stats / tests done for latency, RSS:

AMP fuzzy index — perf & memory (before/after)

Measured on real US-desktop MARS (3,959 suggestions). Before = v0.3.1 (no fuzzy), after = v0.3.2 (with fuzzy).

Build time (per index — runs once per MARS cron resync)

	median	min	max
0.3.1 (before)	7.7 ms	7.4 ms	13.5 ms
0.3.2 (after)	17.3 ms	17.0 ms	21.3 ms
Δ fuzzy	+9.6 ms

Query latency (µs)

path	version	mean	p50	p95	p99
exact	0.3.1 (before)	0.34	0.25	0.96	1.46
exact	0.3.2 (after)	0.35	0.25	1.00	1.46
fuzzy	0.3.2 (after)	4.27	2.29	11.7	27.8

Exact path is unchanged before/after; fuzzy is a new fallback, only on an exact miss.

RSS footprint (per built index)

	footprint
0.3.1 (before)	16.2 MB
0.3.2 (after)	25.0 MB
Δ fuzzy	+8.8 MB (~7 MB by peak-RSS method)

Memory-leak check (0.3.2)

scenario	iterations	RSS
rebuild same index (cron resync path)	400×	flat ~85 MB
fuzzy query hammering	300,000×	flat ~85 MB
build + delete distinct indexes	300×	flat ~85 MB

Summary: ~+10 ms build and ~+8 MB per segment index; zero impact on the exact hot path;back; no leaks.