Fix enumerate() start handling for None and __index__ arguments

[jseop-lim]

Description

Fixes #1074

enumerate() mishandles its start argument in two independent ways. An explicit start=None is treated as if the argument had been omitted, so it silently means 0. And a start that is not already an int is rejected instead of being coerced, so bool and every object implementing __index__ raise TypeError.

AS-IS

>>> enumerate("a", None)
<enumerate object at 0x2671a8f>
>>> list(enumerate("abc", True))
TypeError: 'bool' object cannot be interpreted as an integer
>>> list(enumerate([9, 8, 7], Three()))          # Three.__index__ returns 3
TypeError: 'Three' object cannot be interpreted as an integer

TO-BE

>>> enumerate("a", None)
TypeError: 'NoneType' object cannot be interpreted as an integer
>>> list(enumerate("abc", True))
[(1, 'a'), (2, 'b'), (3, 'c')]
>>> list(enumerate([9, 8, 7], Three()))
[(3, 9), (4, 8), (5, 7)]

(matching CPython)

The two halves ship together because the specializations are shared across call sites. An omitted start arrives as PNone.NO_VALUE, which is neither an integer index nor distinguishable from PNone.NONE by Java type, so guarding one specialization without the other moves the failure rather than removing it:

call sequence in one process	before	default-branch guard only	both guards
enumerate("a", None)	enumerate object	TypeError	TypeError
then enumerate("abc")	[(0, 'a'), (1, 'b'), (2, 'c')]	TypeError — regression	[(0, 'a'), (1, 'b'), (2, 'c')]

Each column is a build. The same branch is already reachable before this change by rejecting a bad start first — enumerate("abc", "x") in a try, then enumerate("abc").

range already coerces its arguments through PyNumberIndexNode and separates an omitted argument with isNoValue guards, so this brings enumerate to the same shape.

Changes

• Guard the specialization that defaults start to 0 with isNoValue(keywordArg) so it matches only an omitted argument, and rename it doNoValue to match.
• Replace the rejecting specialization with doGeneric, which coerces start through PyNumberIndexNode and CastToJavaLongExactNode before the iterator is acquired. Its failure path raises the same ErrorMessages.OBJ_CANNOT_BE_INTERPRETED_AS_INTEGER as before, so None and float keep their exact message.
• Add !isNoValue(start) to that specialization so an omitted argument can never reach it.
• Add tests/test_enumerate_start.py covering the explicit None, bool, an __index__ implementer, one whose __index__ overflows a long, a rejected float, and an omitted start following each of the rejected and coerced cases.

Testing

mx build on linux-aarch64, then mx graalpytest test_enumerate_start.py test_enumerate.py test_list.py test_tuple.py test_iterator.py — 243 tests, all pass. CPython's own test.test_enumerate from lib-python passes as well (92 tests, 14 skipped).

Twelve call shapes were compared against CPython 3.12.10 on the built binary, including the sequences above that exercise the shared specialization state and an __index__ that returns 2 ** 70; all twelve agree. The table above was measured by building the unpatched tree, the one-guard variant, and the full change, and running the sequence against each.