chore: Add more benchmarks

[Abhijeet Prasad (AbhiPrasad)]

Adding a few more benchmarks so I can burn anthropic tokens on pi auto research over night.

[Cedric / ViaDézo1er (viadezo1er)]

Not exactly this PR, but the previous performance improvement PR added sped up bt_json.py. However if a list subclass is given as argument this new fast version doesn't behave as expected:

import dataclasses, sys, os
from braintrust.bt_json import bt_safe_deep_copy

class TaggedList(list): pass

@dataclasses.dataclass
class Point: x: float; y: float

# Plain list: dataclass → dict. TaggedList: dataclass → str (regression).
print(bt_safe_deep_copy([Point(1, 2)]))            # [{'x': 1, 'y': 2}]  ✓
print(bt_safe_deep_copy(TaggedList([Point(1, 2)])))# ['Point(x=1, y=2)'] ✗

And it behaves correctly before the performance improvement change.
I think these functions are public and could be used by users?

-        elif tv is list or tv is tuple:
+        elif tv is list or tv is tuple or isinstance(v, (list, tuple)):

[Abhijeet Prasad (AbhiPrasad)]

Not exactly this PR, but the previous performance improvement PR added sped up bt_json.py. However if a list subclass is given as argument this new fast version doesn't behave as expected:

yeah that's a regression. Feel free to open a gh issue, we can fix as a follow up

I think these functions are public and could be used by users?

nope, we set package_dir={"": "src"}, in setup.py so it shouldn't be exposed. We do expose our test files, but that is a separate issue.

[Cedric / ViaDézo1er (viadezo1er)]

Move the function definition outside the function body so function creating time isn't counted by pyperf?

[Cedric / ViaDézo1er (viadezo1er)]

Anyway good to merge this is just nitpicking