doltliteParseRecordStrict memsets 16 KB per record parse, on the per-row path of every VC read surface

Found in a full-repo review at `0ba280f06f`.

`doltliteParseRecordStrict` zeroes a 16 KB struct on every call, on the per-row path of every version-control read surface.

`src/doltlite_record.c:529`:

```c
memset(pInfo, 0, sizeof(*pInfo));
```

`DoltliteRecordInfo` (`src/prolly_record.h:83-88`) is `4 + 2 × DOLTLITE_MAX_RECORD_FIELDS × 4` bytes, and `DOLTLITE_MAX_RECORD_FIELDS` is `SQLITE_MAX_COLUMN` (2000), so **16,004 bytes** — regardless of how many fields the record actually has. Only `[0..nField)` is ever read, and every reader guards on `nField`.

This is not an incidental path. `diffRecordsEqualFieldwise` (`src/prolly_diff.c:147-210`) declares *two* of these and parses both per row-pair, and it is the equality test behind `prollyValuesEqual`, which diff, merge, blame and status all use per row.

## Measured

Replacing the memset with `pInfo->nField = 0;` in a scratch build, 1M-row table with every row changed, median of 3:

| workload | master | one-line patch | speedup |
|---|---|---|---|
| `count(*) FROM dolt_diff_t` | 1.461 s | 1.073 s | **1.36x** |
| `count(*) FROM dolt_blame_t` | 1.075 s | 0.693 s | **1.55x** |
| `dolt_diff_stat('HEAD~1','HEAD')` | 0.025 s | 0.025 s | 1.00x |
| `dolt_diff_summary('HEAD~1','HEAD')` | 0.026 s | 0.026 s | 1.00x |

Results identical in both builds.

## Fix

Set `nField = 0` instead of zeroing the whole struct — the readers already bound themselves by `nField`, so nothing else depends on the zeroing. Worth doing alongside: shrink `DoltliteRecordInfo` to a small inline array with heap spill for wide records. It is a 16 KB stack object, so it is also the direct cause of most of the oversized frames in the wasm stack report.

Start at doltliteParseRecordStrict in src/doltlite_record.c:529 and inspect DoltliteRecordInfo in src/prolly_record.h:83-88, then trace its two callers in diffRecordsEqualFieldwise at src/prolly_diff.c:147-210. Validate the change against diff and blame workloads, checking that results remain identical and the per-row performance improves without affecting the summary operations.

Written by the indexing model from the issue text.