object_create benchmark: close the last 1ms gap vs Node

[proggeramlug]

Description

Perry scores 9ms vs Node's 8ms on the object_create benchmark (1M new Point(x,y) allocations). This is the only benchmark where Node leads.

The inline bump allocator (v0.5.2) brought this from 318ms to 9ms. The remaining gap is structural:

• Perry's arena allocator has an 8-byte GcHeader per object
• V8 uses hidden classes with direct-mapped field access

Possible approaches:

• Profile to find the remaining per-alloc overhead
• Experiment with smaller GcHeader (4 bytes? type+flags only)
• Batch field initialization with wider stores (write 2 fields per i128 store)

[proggeramlug]

This has been resolved. Perry now scores 2ms vs Node's 14ms on this benchmark (7x faster). The scalar replacement of non-escaping objects (v0.5.17) decomposes new Point(x, y) into stack allocas that LLVM promotes to registers — zero heap allocation. The original issue was about closing a 9ms vs 8ms gap; Perry now leads by a wide margin.