Fix out-of-bounds write in Newton-Raphson division, Improve normalize

[tompng]

VpSetZero, VpSetInf and VpSetNaN left the exponent of the previous value.
VpNmlz wasn'y normalizing the exponent.
Zero, Infinity and NaN have no meaningful exponent and every reader already
ignores it, but the stale exponent reached slice_copy in Newton-Raphson
division. When a quotient block is zero, BigDecimal_fix of the tiny product
x * inv produces a zero that still has a negative exponent. slice_copy
then underflows the copy length and writes one word outside of the quotient
buffer, corrupting the heap.

Reproduction (with the default threshold):

y = 10**1000 + 7
BigDecimal(y * 10**1800 + 5).divmod(BigDecimal(y))

The result is correct, but GC can crash later. With
NEWTON_RAPHSON_DIVISION_THRESHOLD=1, rake test segfaulted in
test_bigmath about once in four runs.

Changes:

• Reset exponent in VpSetZero, VpSetInf and VpSetNaN. This alone
  fixes the crash.
• Fix VpNmlz to normalize exponent
• Rewrite slice_copy to copy only the intersection of the requested range
  and the existing words, so it can never write outside dest[0, length).
• Add a regression test for the zero quotient block.