feat: install host globals natively in the runtime (#38)

[schplitt]

Closes #38.

What

Globals are no longer installed by prepending generated JavaScript to the user's code. The client sends each global as a structured GlobalDef and the Rust runtime installs it natively on the sandbox global object.

globals on the wire changes from List<String> (names) to List<GlobalDef>, a tagged union:

kind	tail	install
bridge	—	bridge stub under name (unchanged)
string	expr	evaluate (expr) as its own script → object.set(name, …)
data (new)	WireValue	wire_to_v8_value → object.set(name, …)
shim	shim, handlerName	stub under handlerName + wrapper built by a fixed factory → object.set(name, …)

Why this closes #38

• Correct stack-trace line numbers — nothing is prepended, so user code starts at line 1. String/shim expressions compile as their own scripts with their own filenames.
• No name interpolation — a global's name only ever travels as a string passed to object.set (or a WireValue field), never pasted into an identifier position. This removes the injection class flagged in the issue's "Additional findings".
• New data-valued global ({ kind: 'data', value }) — constants cross as a WireValue and materialise natively (same supported set as host-module data leaves), killing the JSON.stringify-into-code pattern.

Hot path unchanged

For precompiled prefixes, string/data globals and shim wrappers are baked into the snapshot at precompile() (via install_value_globals in the snapshot isolate); prefix.run() only re-installs bridge stubs, exactly as before. No new per-run cost on the repeated-run path.

Tests

• Rust: 214 passing (added coverage for parsing all four GlobalDef kinds).
• TS: 460 passing — new integration tests for the data global (object, Uint8Array/bigint/null) and a line-number regression test asserting a user error reports line 1 with globals configured.
• cargo fmt clean; clippy introduces zero new lints vs main; eslint/tsc clean.

Notes

• Kept all four global kinds (per discussion on Install globals natively in Rust instead of prepending generated preamble code #38); shrinking the globals surface to modules-only is a possible separate follow-up.
• Wire-protocol change: @iso4/sandbox and the @iso4/v8-* binaries must release together (changeset marks all as minor, matching the fixed group).