Fix progress bar iteration speed accuracy and speed

[PlagueKind]

This change replaces tqdm’s internal rate estimation with a deterministic, time-based EMA smoothing approach.

Problem:

• tqdm’s built-in rate estimator becomes unstable under non-uniform step timing
• save/sample/log pauses cause delayed convergence and “catch-up” behavior in reported it/s

Solution:

• Introduce external time-based rate calculation using perf_counter()
• Apply lightweight EMA smoothing to stabilize display without lag
• Preserve existing postfix metrics system (losses, lr, etc.)
• Avoid modifying training loop behavior, only affects progress display

Result:

• Stable iteration speed reporting
• No catch-up artifacts after pauses
• No dependency on tqdm internal smoothing model

[PlagueKind]

Switched to simplified code by @fatalis