Flux.testmode!() should disable gradient tracking

[staticfloat]

As discussed on slack a while back, the typical Flux evaluation incantation of:

testmode!(model, true)
y_hat = mapleaves(Flux.Tracker.data, model)(feature_mat)
testmode!(model, false)

Should probably be shortened somehow. We can fold the mapleaves() call into testmode!() perhaps by having testmode!() return the baked model, and telling the user to use that new baked model instead.

[MikeInnes]

You could shorten this a bit to:

predict = mapleaves(Flux.Tracker.data, model)
testmode!(predict, true)

And then you can call predict(x) wherever you need to – it also stays tied to the original model, so you can save the mapleaves call. Even the mapleaves will go away with #86 since models will be non-tracking by default.

Until then I'd be happy to have a simple out-of-place testmode function that does both things at once.

[MikeInnes]

Pleasingly, with #643 this all becomes a no-op.