platforms: cache results or reduce overheads

[oliver-sanders]

platform_from_name is fairly quick, ~1e-5 for me, however, it can get called a lot of times causing this to add up substantially.

Consider this example:

#!Jinja2

{% set N = 1000 %}

[scheduling]
    [[graph]]
        R1 = """
{% for i in range(N) %}
            x_{{ i }}
{% endfor %}
        """

[runtime]
{% for i in range(N) %}
    [[x_{{ i }}]]
{% endfor %}

There is only one platform, however, platform_from_name will be called at least 1000 times during validation.

Here's profiling results from a large parameterised workflow showing ~20s spent on platform_from_name during validation:

This time is almost all spent on deepcopy and regex stuff.

Pull requests welcome!