feat: bytes read and bytes write mectirc for per lance scanner and writer

[SaintBacchus]

When we scan the lance dataset, it's better to provide the read metrics, especially for the io bytes and IOPS metric.

The Spark web UI has some task-level metrics, for example, the input bytes. But now Lance can't provide it.

The same metrics are required when we write a lance dataset.

[westonpace]

#3575 adds these metrics on the read side. We might need to add some new methods for accessing these metrics. Currently they are logged.

[Joaco2603]

Hi team! I'd like to work on this issue to implement I/O metrics collection for Lance dataset operations.

My Implementation Plan:

Reading Metrics (Scanning):

• Track bytes read from data files during scan operations
• Count read operations for IOPS calculation
• Integrate with existing scan execution flow

Writing Metrics (Saving):

• Monitor bytes written during dataset saves
• Track write operations for IOPS
• Hook into the write path without affecting performance

Approach:

• Create IOMetrics struct to encapsulate the metrics
• Use atomic counters for thread-safe operations
• Add metrics collection to key I/O operations
• Expose metrics through existing profiling interfaces

I have experience with Rust and system metrics collection. Looking forward to contributing!