perf(vm): reuse linked Atoms for static property names

[Eric-Song-Nop]

Stacked on #11. Base: perf/array-own-key-early-check; head: perf/static-property-atoms, both in pocket-stack/quickjs-oxide. Review and merge in dependency order.

Migrated from Eric-Song-Nop#2 with identical commits and preserved measurement/validation evidence.

修改 / Change

Static-name operations used to fetch a String constant and intern it on every execution. Link those names to runtime-local Atoms once at publication and reuse a constant-indexed map in normal calls and resumed activations. Preserve String constants, serialized representation and existing PropertyKey reference management. Bytecode owns the new Atom references through its existing transaction/cleanup path; functions without static names allocate no map.

静态名称指令此前每次执行都取字符串并重新 intern。现在发布时链接一次，普通调用和恢复执行的帧按常量索引复用 Atom。保留字符串常量、序列化形式和现有 PropertyKey 引用管理；由字节码的现有事务与清理路径持有/释放 Atom，无静态名称的函数不分配映射。

Related issue / 关联 issue: #7

实测 / Measurements

Same release flags/machine, diagnostics off, five alternating before/after repetitions; all 40 runs passed output checks. / 同一机器和 release 参数、关闭诊断，五轮交替对照，40 次运行输出校验全部通过。

Workload / 工作量	Before median ms	After median ms	Time reduction / 耗时下降
Fixed property reads ×20 million / 属性读取 2,000 万次	10813.374	7240.065	33.05%
Fixed Richards ×10 / 固定 Richards 十次	2117.465	1477.229	30.24%
Original Richards harness / 原始 harness	8460.681	6163.680	27.15%

Original Richards median score: 16.7 → 24.1 (+44.31%). Whole-process times include initialization/compilation/teardown and are distinct from harness scores. Empty-loop control ranges overlap; no numeric-loop improvement is claimed.

原始 Richards 中位分数 16.7 → 24.1（+44.31%）。完整进程耗时包含初始化、编译和收尾，与 harness 分数是不同指标；空循环对照范围重叠，不据此声称数值循环优化。

Compile/publication medians: property reads 0.094 → 0.101 ms, original Richards 12.697 → 12.586 ms. Original Richards retains 6,912 map bytes + 5,824 auxiliary-reference bytes, excluding allocator/Rc headers, Atom-table/string storage and VM-frame changes; live table-backed Atoms increase 471 → 518 in the compilation snapshot. These are measured costs, not total memory overhead or a general speedup guarantee.

编译加发布中位时间：属性读取 0.094 → 0.101 ms，原始 Richards 12.697 → 12.586 ms。原始 Richards 新增 6,912 字节映射 + 5,824 字节引用数组载荷，不含 allocator/Rc 头、Atom 表/字符串和 VM 帧变化；编译快照中的表内 Atom 471 → 518。这不是总内存增量，也不是普遍收益保证。

双语报告 / Report · 全部样本、哈希及对照 / Samples, hashes and comparison

验证 / Validation

• 2,998 Rust tests passed, 0 failed, one existing ignored; ownership/rollback, runtime isolation, numeric/UTF-16 keys, getter/Proxy/prototype behavior, resumed generators and malformed mappings covered. / 2998 项通过，零失败，1 项原有忽略，覆盖上述语义及生命周期。
• All 102,037 Test262 outcome rows match the parent revision, zero changed rows/new failures. The frozen full-file receipt still fails because the source fingerprint changes the header; no baseline updated. / 完整 Test262 前后逐行一致；冻结完整文件校验仍因源码指纹变化失败，没有更新基线。
• Formatting, source layout, Rust-only and strict production Clippy passed (Rust 1.88; profiling, default and test262-host configurations). / 格式、布局、Rust-only 和三种配置的生产目标 strict Clippy 通过。
• Binary-object production boundary passed remotely; all 690 isolation canaries were rejected. / 远端字节码生产边界检查通过，690 个隔离 canary 全部被拒绝。
• All 480 compilation-probe iterations retained, with 432 non-warmup samples. CPU stacks were not re-sampled in this round. / 保留全部 480 个编译探针迭代，其中 432 个非 warmup；本轮未重新采集 CPU 栈。