bench: add Polars streaming vs in-memory loader benchmark

[AhmedAli58]

What this does

Adds a benchmarking script that compares PyHealth's new Polars streaming data loader
against the legacy in-memory loader across RAM usage, wall-clock time, and throughput
at 3 dataset scales (100, 1k, 5k patients).

Why it matters

The new streaming loader was added without systematic benchmarks. This PR gives
maintainers and users data to make informed decisions about which loader to use
based on dataset size and available memory.

How to run

python benchmarks/loader_benchmark.py

Results

See benchmarks/results.csv and benchmarks/benchmark_chart.png for outputs.

[Logiquo]

InMemorySampleDataset is primarly designed for unittest, it is not as battle tested as the SampleDataset.

And also I think InMemorySampleDataset lacks a few functionality compare with SampleDataset.

[jhnwu3]

Hey we did run some tests across a variety of things to make our decision: https://pyhealth.readthedocs.io/en/latest/why_pyhealth.html

Will be closing this PR. Would love to talk more if you want to discuss benchmarking PyHealth 2.0 in general on the discord