Fix a couple of inferrability issues

[timholy]

Discovered while kicking the tires on Unitful. I'm not sure whether there's a concern about making tol a positional argument for rationalize, but the solution I adopted seems reasonable to me. Even when you rely on the keyword argument for tol, this still makes the function faster:

Julia 0.5:

julia> @benchmark rationalize(3.2)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     195
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  224.00 bytes
  allocs estimate:  5
  minimum time:     494.00 ns (0.00% GC)
  median time:      499.00 ns (0.00% GC)
  mean time:        556.76 ns (7.63% GC)
  maximum time:     31.35 μs (97.69% GC)

This branch:

julia> @benchmark rationalize(3.2)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     525
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  192.00 bytes
  allocs estimate:  3
  minimum time:     213.00 ns (0.00% GC)
  median time:      215.00 ns (0.00% GC)
  mean time:        233.26 ns (3.46% GC)
  maximum time:     2.98 μs (87.50% GC)

and

julia> @benchmark rationalize(Int, 3.2, $(eps(3.2)))
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     861
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  0.00 bytes
  allocs estimate:  0
  minimum time:     140.00 ns (0.00% GC)
  median time:      141.00 ns (0.00% GC)
  mean time:        146.25 ns (0.00% GC)
  maximum time:     319.00 ns (0.00% GC)

[tkelman]

waiting until 0.5.1