Improve performance

We are still a bit behind better-sqlite3. For example, there's more overhead in get():

penberg@vonneumann libsql-js % node perf/perf-libsql.js

cpu: Apple M1
runtime: node v18.19.0 (arm64-darwin)

benchmark      time (avg)             (min … max)       p75       p99      p999
------------------------------------------------- -----------------------------
• Statement
------------------------------------------------- -----------------------------
get(1)      1'969 ns/iter   (1'911 ns … 2'539 ns)  1'970 ns  2'424 ns  2'539 ns

summary for Statement
  get(1)

penberg@vonneumann libsql-js % node perf/perf-better-sqlite3.js
cpu: Apple M1
runtime: node v18.19.0 (arm64-darwin)

benchmark      time (avg)             (min … max)       p75       p99      p999
------------------------------------------------- -----------------------------
• Statement
------------------------------------------------- -----------------------------
get(1)        587 ns/iter     (555 ns … 1'122 ns)    584 ns    944 ns  1'122 ns

The assumption that a lot of this is memory memory allocations and copies as values are passed between the C, Rust, and JavaScirpt layers. We also suspect that Neon Bindings may be less efficient than better-sqlite3 is which seems to hook into the V8 APIs directly.

Let's improve the benchmarks, profile, and fix performance to be on-par with better-sqlite3.

Start by running perf/perf-libsql.js and perf/perf-better-sqlite3.js to reproduce the reported get(1) gap, then profile the benchmark across the C, Rust, and JavaScript layers. Done means improved benchmarks and performance on par with better-sqlite3, with the relevant profiling results supporting the change.