Clarify lazy SST opening in turbo-persistence docs (#100311)
## What?
Fixes how the turbo-persistence README describes opening the database
and reading from it. The old text said every SST file is memory mapped
up front.
## Why?
The old wording overstated how much memory opening a database costs. In
the current implementation:
- Opening the database loads only the meta files (memory mapped, or read
into heap memory in `AccessMode::File`). It parses their AMQFs and
filters out obsolete entries. SST and blob files are not opened.
- An SST file is opened lazily, the first time a lookup gets past its
hash-range and AMQF checks, and is then cached on its meta entry.
- `AccessMode::File` never memory maps SST files. It keeps the block
offset table on the heap and reads blocks with positioned reads.
- In either mode, opening an SST reads and verifies only its index
block. Key and value blocks are read when a lookup needs them.
## How?
- Updates the **Opening** steps to match the directory-loading code.
- Adds a short lifecycle explanation to **Reading**, covering both
access modes and what drives memory cost.
- Corrects the lookup outline:
- adds the hash-range check
- says the AMQF lives in the meta file
- shows where the SST is opened
Documentation only. The wording was checked against
`Db::load_directory`, `MetaEntry::sst`, `MetaFile::lookup_entry` and
`StaticSortedFile::open`.
## Verification
- Stacked on #100310, which removes the README entry from
`.prettierignore` and formats all READMEs. `npx prettier@3.6.2 --check
turbopack/crates/turbo-persistence/README.md` passes.
- `git diff --check`
<!-- NEXT_JS_LLM -->
<!-- fleet 9072af3c-95f7-4175-a93d-5b232fe92150 -->
Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Niklas Mischkulnig <4586894+mischnic@users.noreply.github.com>
