Fix over-segmentation in BlackRock

[h-mayorquin]

This will fix #1770

Testing data is merged here https://gin.g-node.org/NeuralEnsemble/ephy_testing_data/pulls/167 and is used on testing here.

This PR fixes #1770 using the methodology discussed on #1773. To do this, we do three things:

1. We separate the parsing of the data blocks from the segmentation of the data
2. I also continue the refactor started in Blackrock: improve nev data reading #1772 and Blackrock: improve nev header reading #1771 to remove dynamically loaded functions per spec. Now, the data types of the headers and the data blocks are global variables and are separated from the code.
3. The gaps report (see Blackrock add summary of automatic data segmentation  #1769) was only working for the PTP format. All the formats with timestamps report gaps now.
4. Following the discussion on General API for handling sample gaps on rawio #1773 this PR introduces the new gap_tolerance_ms to give user control over the size of the gaps that should create segments. By default, it is None and an error is raised if gaps are found. User can then opt-in to load the data by specifying a gap size.
5. I used a buffered version of the memmaps so we don't need to create a memamp per block. This reduces the number of memmaps created by the reader which is an OS limit.

This is a large PR but I think that @samuelgarcia prefers them that way rather than sausage sliced PRs and he is the person that might end up reviewing it. I am happy to break it apart if whoever is gona review this prefers it another way.

Tagging @cboulay here as requested in case if he has time to check it.

[h-mayorquin]

Here, we use views on a linear memmaped buffer to avoid creating a memmap per data block. This is good because OSes usually limit the ammount of memamps you can create.

[h-mayorquin]

Changed here to do this after segmenting.

[h-mayorquin]

This is the core of the PR:

1. We parse the data blocks (I improved the docstrings there)
2. We segment the file and report if necessary
3. We transform back to the previous data structures to keep the diff small.

[h-mayorquin]

Here is the same technique to avoid many memmaps: create an array with the buffer (one memmap) and then create views into it.