[Feature Request] Enable and Disable profiling for warm up

### Describe the feature request

I would like to do some profiling with `onnxruntime` using the utilities described here [here](https://onnxruntime.ai/docs/api/python/auto_examples/plot_profiling.html).
But following this tutorial, I can only set `session_options.enable_profiling=True` before the inference session starts and end the profiling with `session.end_profiling()` once and for all.

I want to be able to:
- warm up the session/model before the profiling starts.
- compute other kind other statistics like `stddev` of a kernel/node's runtime.

This can be achieved by having a different profiling file each time `session.end_profiling()` is called like in this example:

```python
from optimum.onnxruntime import ORTModelForSequenceClassification
from tempfile import TemporaryDirectory
import onnxruntime as ort
import pandas as pd
import numpy

inputs = {
'input_ids': numpy.array([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]]),
'attention_mask': numpy.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]),
'token_type_ids': numpy.array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]),
}

with TemporaryDirectory() as tmpdirname:
session_options = ort.SessionOptions()
session_options.enable_profiling = True
session_options.profile_file_prefix = tmpdirname + '/profiling'
model = ORTModelForSequenceClassification.from_pretrained(
"bert-base-uncased",
export=True,
session_options=session_options,
)

for _ in range(5):
model(**inputs)

warmup_file = model.model.end_profiling()
# after this maybe a new profiling file can be created

for _ in range(10):
model(**inputs)

profiling_file = model.model.end_profiling()
profiling_df = pd.read_json(profiling_file, orient='records')
# processing profiling_df
```

### Describe scenario use case

stable and in-depth profiling with no outliers.

Start with the Python profiling example linked in the issue and the SessionOptions.enable_profiling and session.end_profiling APIs. Trace their implementation and existing profiling tests, if any; done should include a defined way to warm up before profiling and produce separate profiling data for later runs.

Written by the indexing model from the issue text.