Eps as minimum scale

[quic-kyunggeu]

Problem

Currently v2 encoding analyzer asserts scale >= torch.finfo(...).tiny.
However, this tiny is the smallest positive normal number, which is numerically extremely unstable

Fix

Use torch.finfo.eps as minimum scale instead.
eps is a much more moderate value than tiny, and hence more numerically stable

	float32	float16	bfloat16	float64
tiny	1.18e-38	6.10e-05	1.18e-38	2.23e-308
eps	1.19e-07	9.77e-04	7.81e-03	2.22e-16

(p.s. 1.0 / (2**16-1) is equal to 1.53e-05, which is still larger than float32 eps)