Update to upstream version '6.5.0'
with Debian dir 58d5f805181a3d5452f6da0fec2d4dd6cec629b6

This reduce the number of function call and should be faster.

The mashup of blind optimisation in the previous changeset yield some
interesting results in total.

It would be insightful to measure them individually, but that would
take more time than we currently have.

When testing all the validator changes on our previous "benchmark" we
see quite interesting improvement.

    swh loader run mercurial https://foss.heptapod.net/mercurial/mercurial-devel directory=/data/repos/mercurial-devel

= Median time of 3 run =
base:   17 minutes 48 seconds
before: 11 minutes 50 seconds
after:  11 minutes 11 seconds

On a profile of the same run, the `to_model` call of the from_disk's `Directory` class took the following percentage:
base:   43%
before: 15%
after:  11%

(This commit is actually doing two things /o\)

- we inline the type-checking in the custom validator to reduce the
  number of function call.

- we optimize some of the custom validator by skipping the creation of
  intermediate tuples.

Since all `generic_type_validator` are optimized away, the code will no
longer be called. So we remove that code to avoid any drifting.

A nice "exception" is provided in case this start getting called again
in the future.

Since try/except context are known to be expensive in Python, it seems
useful to remove them.

This ensure we don't have any remaining `generic_type_validator` call
that have not been optimized away.

This indirection seems useless and is probably the remains of some long
forgotten rituals.

This should reduces function calls and speeds things up.

It might be useful to introduce even more specialized validator in the
future. It would also be useful to skip the intermediate try/except.

Some of this will be done in later changesets.

This is currently doing nothing, but prepare for actually changing the
generic validator into faster specialized variants.

Before this change we would do the following :

1) translate from_disk's object into `dict`,
2) sort these dict,
3) feed the list to `Directory.from_dict`,
4) create DirectoryEntry from these dict.

Skipping the directory creating and directly creating the
DirectoryEntries provide us with a small but stable and noticeable
performance win.

We tested this change on simple information of the Mercurial loader,
with a noop-loader stockage:

    swh loader run mercurial https://foss.heptapod.net/mercurial/mercurial-devel directory=/data/repos/mercurial-devel

= Median time of 3 run =
before: 11 minute  56 seconds
aftere: 11 minute  50 seconds

On a profile of the same run, the `to_model` call of the from_disk's `Directory` class took the following percentage:
before: 17%
after:  15%

Do not create model object while sorting entry before creating model
object.

This is another case of "let us create object X to prepare the creation
of object X", slowing things down.

In practice, we will likely skip this code-path after the next
changeset, however this seems useful to get this performance footgun
out the way.

We tested this change on simple information of the Mercurial loader,
with a noop-loader stockage:

    swh loader run mercurial https://foss.heptapod.net/mercurial/mercurial-devel directory=/data/repos/mercurial-devel

= Median time of 3 run =
before  12 minutes 59 seconds
after:  11 minute  56 seconds

On a profile of the same run, the `to_model` call of the from_disk's `Directory` class took the following percentage:
before: 24%
after:  17%

Before this change, a Directory object was built to compute the `id` of
we fed to the Directory object we built for `to_model`.

We tested this change on simple information of the Mercurial loader,
with a noop-loader stockage:

    swh loader run mercurial https://foss.heptapod.net/mercurial/mercurial-devel directory=/data/repos/mercurial-devel

= Median time of 3 run =
before: 17 minutes 48 seconds
after:  12 minutes 59 seconds

On a profile of the same run, the `to_model` call of the from_disk's `Directory` class took the following percentage:
before: 43%
after:  24%

Update to upstream version '6.4.1'
with Debian dir 057ba180164d4c2d0d5dff8360fd67d0f915cc5b

Format was changed in 752fb81d.

Update to upstream version '6.4.0'
with Debian dir 57a9924f627e97fd6dad7f13a16705030ad01fca

This is needed by swh-scrubber when recomputing the hash of such snapshots.

It allows building a snapshot manifest despite some branches being
unresolved (instead of raising an error).

This feature was removed in 57ae405d
but in the end, Snapshot.compute_hash() will need to use it so that
swh-scrubber does not crash when checksum such snapshots.

Update to upstream version '6.3.1'
with Debian dir 9bbf55bd07c86e39151c728804d2814904ab62d9

hypothesis.strategies.integers() is inclusive, but microseconds values
greater or equal to 1 million are rejected by the model