tp: cut allocation and string work in the executor and CREATE TABLE (#7706)

**Summary**

Removes avoidable allocation and string work in code that runs whenever
an operator retains rows, an `INTERVAL INTERSECTION` runs, or `CREATE
PERFETTO TABLE` materializes a query. Each change is small; together
they are 5-67% faster on the paths they touch.

**Changes**

**RowStore**
- Reuses the chunks it copies borrowed columns into. It used to allocate
a new chunk for every column of every batch and free them all on
`Clear()`, which dominated operators run over and over on small inputs.
A chunk is only reused once nothing else holds it.
- Copies validity a word at a time (new `BitVector::FillBits`) instead
of a bit at a time.

**IntervalIntersect**
- Groups lived in a `FlatHashMapV1`, whose minimum capacity of 1024
slots was walked and cleared on every rerun, and each group's intervals
were freed each time. Groups are now reused across runs, held by pointer
so their keys never move, and found through a `FlatHashMapV2` of
`string_view`s.

**Strings in `CREATE PERFETTO TABLE`**
- The dataframe virtual table returned strings without their length, so
SQLite ran `strlen` on every one. It now passes the length and marks
them NUL-terminated (`SQLITE_UTF8_ZT`).
- Strings read back from the string pool were hashed and interned again.
A pointer into the pool is now mapped straight back to its id, checked
against the size prefix before it and the terminator after it. This is
done under the pool's lock, which is only taken when locking is enabled.

**Benchmarks**

Functions 64-byte aligned in both builds, so code placement does not
skew results; best of 5.

| Benchmark | Before | After | Ratio |
|---|--:|--:|--:|
| `BM_RowStore` (16 batches) | 216 µs | 92 µs | 0.43x |
| `BM_RowStoreOneRow` | 201 ns | 66 ns | 0.33x |
| `BM_IntervalIntersect` (100K rows) | 5.00 ms | 4.09 ms | 0.82x |
| `BM_IntervalIntersectFourRows` | 3.27 µs | 0.50 µs | 0.15x |

Real traces, query time:

| Trace | `CREATE PERFETTO TABLE` copies of sched, thread_state, slice,
counter, thread | `INTERVAL INTERSECTION` of sched and thread_state `PER
utid`, 5 times |
|---|--:|--:|
| `wattson_syscore_suspend` | 122 → 111 ms (0.91x) | 117 → 116 ms
(0.99x) |
| `chrome_android_systrace` | 126 → 112 ms (0.89x) | 120 → 109 ms
(0.91x) |
| `example_android_trace_30s` | 77 → 74 ms (0.96x) | 230 → 222 ms
(0.97x) |
| `android_sched_and_ps` | 40 → 38 ms (0.95x) | 124 → 114 ms (0.92x) |

Results are identical to before on all four traces.

**Testing**

New unit tests for `BitVector::FillBits` and for `RowStore` reusing
chunks without overwriting ones still held. Unit and diff tests pass.