Added tag 0.20.0 for changeset b055388f78a3

changelog: document the update `keepnans` keyword

api/update: fix the documentation

migrate/0.20: add command to migrate diffstart/end data for one series

cli: add checkdiff-random command

search: fix signature of by.source

Deriving these things automatically would be a good idea.

api/rename: add a `propagate` parameter

This will be used by extenders (notably formula).
In the future, all of this will be moved where it belongs.

migrate/0.20: make the data migration deferrable to a task

appease the linter

series.sql: make diffstart/diffend optional

We also show the impact on the `insertion_dates` call with
from/to_value_date restrictions.

migrate/0.20: more cleanups

util/diffs: factor a common implementation for diffs

It is now used by tsio.diffs and the 0.20 migration.

api/update: make erasure update explicit

We introduce a `keepnans` keyword whose default value is False.

By default, series given to .update will be stripped of their Nans,
unless we explicitly tell it it is for point erasure purposes.

It is not only for erasure: nans won't be rejected if they apply to
previously unvalued points.

Why would one want "pure" nans ?  A series can also be used as a patch
for another series (as in the formula `priority` operator) and nans
can make sense in this context (they indeed will erase data).

Because of these changes, in series.sql, tsstart and tsend can be null.

test/idates: show the behaviour of insertion_dates with nans

Including a revision which contains only nans and does not erase anything.

util/diff and tsio/create: regularize the nan values handling

They were special cased in `diff` and tsio/create, in wrong ways.

There's nothing special about a nan, it's a floating point value.
The erase semantics does not change anything.

At creation time, they seem to be useless. But what if we create a
series whose purpose is to patch another ? The formula system
certainly encourages that kind of thing. So we take the nans and we
don't touch them.

test/tsio: exhibit odd behaviour with the erasure of non existent data

test/util/diff: exhibit issue with diff computing involving nans

pkg: numpy 2 is out, let us not use it !