Add performance analysis to hvd

[zilbermanor]

Added easy performance analysis and tuning capabilities to MLRun's MPIJob runtime.

Horovod Autotune

<mpi-function>.with_autotune(log_file_path: str, warmup_samples: int, steps_per_sample: int, bayes_opt_max_samples: int, 
gaussian_process_noise: float)

Adds an Autotuner to help optimize Horovod's Parameters for better performence. The autotuner will collect metrics and tune horovod's parameters while running using Bayesian optimiation. This may affect the performence of the run initially but after arriving to the best parameters should increase performence.

Since autotuning imposes a tradeoff between early performence for better performence later on, It's advised to enable it when both:

• Training should take a long timeout
• Scaling efficiency was found lacking with the default settings

Horovod Timeline
<mpi-function>.with_tracing(log_file_path: str, enable_cycle_markers: bool)
Add Horovod Timeline activity tracking to the job to analyse its performence. The data will be saved as JSON to {log_file_path}. It can then be viewed via a trace viewer like chrome or edge's edge://tracing.

[Hedingber]

one little typo