RFC/WIP: coercion "cascades" for MvNormal

[timholy]

This redesigns the constructors of MvNormal to:

• check argument consistency in the inner constructor
• support "coercion," for example MvNormal{Float32}(μ, Σ)
  will enforce that the result has eltype Float32
• cascade coercion among the multiple type parameters

The benefits of this design are subtle but significant, as it
allows the user to specify intent, allows the constructors to
accept input arguments that are not (yet) consistent with
the desired output types (e.g., MvNormal{Float64}(Any[1,2], I)),
and makes it easy for developers to "propagate" consistent
types across the code base (e.g., fixes #1544).

Among things this fixes, it's currently possible to construct an MvNormal with inconsistent dimensions:

julia> dist = DiagNormal(zeros(3), PDiagMat(zeros(2)));

julia> rand(dist, 5)
ERROR: DimensionMismatch: Inconsistent argument dimensions.
Stacktrace:

This is possible because the consistency check is not in the inner constructor.

[timholy]

This won't currently pass tests because PDMats needs similar improvements. But I thought this might be worth posting for comment first before going to larger efforts.