Print seed on test failure to aid reproduction

[DRMacIver]

Users often find themselves in one or more of the following scenarios:

• They have data types with very bad reprs
• They are using Hypothesis features that currently don't play well with @example.
• Their CI setup doesn't interact brilliantly with the Hypothesis example database.

This can make it annoying to impossible to reproduce failures from CI locally.

This PR aims to mitigate that: When a test fails, if the random seed required to reproduce that test was not already set deterministically (with @seed or derandomize=True or with the command line argument I'll mention in a moment), Hypothesis will print out the seed it used along with instructions on how to use it to reproduce the example.

In order to make that information more readily useful and allow reproducing examples without editing the source code, this also adds a command line flag for pytest users to pass the seed explicitly.

[Zac-HD]

LGTM - this is the kind of important but unglamorous work that distinguishes really cool projects from really useful projects. Thanks to Smarkets for funding it!

[Zac-HD]

"lead" -> "led" or "leads"; or just reread the whole sentence and decide on a tense.

[Zac-HD]

Not much more to say really - it's a solid and useful change, but probably not something to write a blog post about 😄

[DRMacIver]

Not much more to say really - it's a solid and useful change, but probably not something to write a blog post about 😄

I am in fact scheduled to write a blog post about it. 😉

More specifically, Smarkets are paying for a bunch of feature dev, mostly around making Hypothesis performance problems more legible (this fits into that because it lets you reproduce the exact scenario that lead to a failed health check, which I totally forgot to mention as a use case in the preamble above!). Once I've finished all of them I'm going to write a blog post about the whole set of work and the motivations leading up to it.