total_seconds: Fix precision problem

[jepler]

Since CircuitPython floats have very small precision, the "total_seconds" method was not very useful for large timedeltas.

Instead of always returning a float, use pure-integer arithmetic (and return an int) if either:

• the length of time is big (1<<21 is 2097152 seconds or about 24 days)
• the number of microseconds is zero

Otherwise, for small values with nonzero microseconds, use floating-point math.

The cut-off point was chosen because in CircuitPython float arithmetic, 2097151.0+0.5 is different from 2097151.0, but 2097152.0+0.5 and 2097152.0 are the same float.

[brentru]

lgtm - thanks jepler!