Fix profiler

[astachowiczhabana]

This pull request introduces profiling enhancements to the text generation script (examples/text-generation/run_generation.py) by integrating a per_sequence_profiler and improving the handling of profiling during generation. The changes aim to provide more granular performance metrics and better control over profiling behavior.

Profiling Enhancements:

• Added profiler as a parameter to the generate function, allowing the use of either disabled_profiler or per_token_profiler based on the disable_profiling flag. [1] [2]
• Integrated per_sequence_profiler into the generation loop to measure performance at each step, including starting, stepping, and stopping the profiler at appropriate points. [1] [2] [3] [4]
• Modified the generate_dataset function to include a disable_profiling flag, ensuring profiling can be disabled when generating datasets. [1] [2]

Code Refinements:

• Reordered calls to timer.step() and per_sequence_profiler.stop() to ensure accurate timing and profiling data collection.