Add facility to specify a default logger different from `nothing`

[ablaom]

This could be provided as either part of #925 or as a follow-up PR.

The idea is to define a global variable DEFAULT_LOGGER and to make the default value of the logger keyword option in evaluate! equal to DEFAULT_LOGGER, rather than nothing. (And later, do the same for TunedModel and IteratedModel; see https://github.com/alan-turing-institute/MLJ.jl/issues/1029.)

In __init__ the DEFAULT_LOGGER would be set to nothing but MLJBase will provide a public method default_logger to set and inspect the value. For example:

# inspecting:
julia> default_logger()
nothing

# changing:
using MLFlow
julia> default_logger(MLFlowLogger())
MLFlowLogger(MLFlow(
    baseuri = "http://127.0.0.1:5000", 
    apiversion = 2.0
), "MLJ experiment", nothing)

The required code can be modelled exactly on the DEFAULT_RESOURCE global variable (see src/init.jl) and the default_resource setter/getter (see src/MLJBase.jl).

To do after this integration:

• [ ] Have MLJ re-export default_logger

@pebeto

[pebeto]

Does this change replace our keyword argument in evaluate! and save?

[ablaom]

Yes, everywhere we have a keyword argument logger=nothing , we need to change to logger=default_logger(). At present, that's just evaluate!.

In the case of MLJBase.save, there is nothing in MLJBase that connects save to logging. That happens only in MLJFlow.jl, where MLJBase.save(::MLFlowLogger, ::Machine) is overloaded. I suppose we could arrange that MLBase.save(mach) saves the machine to the default logger, ie, do something like this in MLJBase:

const ERR_INVALID_DEFAULT_LOGGER = ArgumentError(
    "`default_logger()` is currently `nothing`. Either specify an explicit path or stream as "
    "target of the save, or use `default_logger(logger)` to change the default logger. "
)
"""
    <new doc string>
"""
MLJBase.save(mach::Machine) = MLJBase(default_logger, mach)
MLJBase.save(::Nothing, ::Machine) = throw(ERR_INVALID_DEFAULT_LOGGER)

What do you think?

[pebeto]

Awesome, how it could be used by MLJTuning and MLJIteration? Are we going to call to certain functions defined by us inside that packages?