Switch raw dataset indexing from `fsspec` to `obstore`

## 🚀 Feature

Switch raw dataset indexing from `fsspec` to `obstore`
> `obstore` is 2 to 3x faster

Initial experiments resulted in some sort of deadlock in streaming. It worked for indexing but not for streaming — only one thing worked at a time.

### Motivation

> `obstore` is 2 to 3x faster for indexing. really useful for large datasets

Follow up to https://github.com/Lightning-AI/litData/pull/652, commit https://github.com/Lightning-AI/litData/pull/652/commits/39c6e6e9ddb52eb449a2cd62290dbb025eb518e8

### Pitch

### Alternatives

### Additional context

Review pull request #652 and commit 39c6e6e9ddb52eb449a2cd62290dbb025eb518e8 first, focusing on the raw dataset indexing and streaming behavior described there. The change is complete when raw dataset indexing uses obstore for the stated performance benefit without reproducing the deadlock that prevented indexing and streaming from working together.

Written by the indexing model from the issue text.