For very out-of-range floating-point constants (1e100000000),
precise formatting of the offending value for error messages
is not needed and potentially extremely slow.

This change resurrects an adjusted variant of the original code
which uses float64 formatting in the common case (in-range values),
and a fast manual approximation for out-of-range values.

Change-Id: I2f6e53040929b8bf924dac4bb27c4d811ede48e2
Reviewed-on: https://go-review.googlesource.com/8470Reviewed-by: Alan Donovan <adonovan@google.com>