feat: add vm.getAllocatedSize() for constant-time memory sampling

Problem

vm.getMemoryUsage() is the only way to read the runtime's memory use, and it's backed by JS_ComputeMemoryUsage, which walks every object, string and shape on the heap. That makes sampling memory on every host call O(heap): a host that tracks a guest's memory as it runs ends up spending most of its CPU in the walk once the guest holds a sizeable dataset.

Measured with a local benchmark script, not included in this PR (Apple M4 Pro, bun, best of 5 rounds):

10,000 samples on the 100 MB heap: ~52 s with getMemoryUsage() (extrapolated from 100 calls), ~0.1 ms with getAllocatedSize().

Change

• New C export qjs_get_allocated_size, which returns the byte counter that the runtime's malloc functions in c/interface.c already maintain on every allocation and free (qjs_mem_used, added in #33).
• New vm.getAllocatedSize(), which reads it in O(1).
• getMemoryUsage() gains an allocatedSize field holding the same value, so the full breakdown shows it too.
• README note under "Memory limits" on when to use which; changeset (minor).

Tests

test/allocated-size.test.ts:

• matches getMemoryUsage().allocatedSize and stays close to mallocSize on a fresh VM, after allocating a large array, and after freeing it and running GC
• grows by its delta in step with mallocSize, and counts ArrayBuffer backing stores
• reaches, but never exceeds, memoryLimit when retained buffers trigger OOM
• after QuickJS.restore() equals the value before the snapshot, and keeps tracking allocations and frees

test/memory-usage.test.ts checks the new field. Full suite: 42 files, 2082 tests passing against a rebuilt quickjs.wasm (wasi-sdk 32).