perf(bt_json): optimize _to_bt_safe and bt_safe_deep_copy hot paths

[Abhijeet Prasad (AbhiPrasad)]

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
  primitive-int         2.02 us ->  42.2 ns   47.8x faster
  primitive-float-nan   2.01 us ->  67.8 ns   29.7x faster
  str-subclass-enum     2.02 us ->  97.5 ns   20.8x faster
  dataclass             10.2 us ->  3.32 us    3.1x faster
  pydantic-v2-like      1.68 us ->  1.78 us    1.1x slower (noise)
  pydantic-v1-like      1.94 us ->   841 ns    2.3x faster

bt_safe_deep_copy:
  small                 19.7 us ->  2.30 us    8.6x faster
  medium                 179 us ->  17.3 us   10.4x faster
  large                1.66 ms  ->   138 us   12.1x faster
  circular               178 us ->  17.9 us    9.9x faster
  non-string-keys       16.7 us ->  2.12 us    7.9x faster

[ViaDézo1er / cedric (viadezo1er)]

Tested on a small Linux machine, the speedup ratio is the same