fix: source unstable-period ids from the C header (#752)

[mario4tier]

Fixes #752.

Bug

_ta_func_unst_ids was built with enumerate() over the TA-Lib 0.4.0 names.
TA-Lib C 0.6.0 inserted TA_FUNC_UNST_IMI at slot 12, shifting every id from
there up. setup.py requires ta-lib >= 0.6.1, so every normal install is
affected: set_unstable_period('RSI', n) sets PLUS_DM's unstable period,
'MAMA' sets KAMA's, and 'ALL' sets T3's alone.

It went unnoticed because set and get look the id up in the same table, so a
wrong id round-trips perfectly.

name	old	new	what the old id hit
ADX … HT_TRENDMODE	0–11	0–11	correct
KAMA	12	13	IMI
MAMA	13	14	KAMA
MFI	14	no-op	MAMA
MINUS_DI	15	16	MFI
MINUS_DM	16	17	MINUS_DI
NATR	17	18	MINUS_DM
PLUS_DI	18	19	NATR
PLUS_DM	19	20	PLUS_DI
RSI	20	21	PLUS_DM
STOCHRSI	21	no-op	RSI
T3	22	23	STOCHRSI slot (never read)
ALL	23	24 / 65535	T3 only
ADXR	1	no-op	ADXR slot (never read)

Fix

Ids now come from lib.TA_FUNC_UNST_*. Cython reads a cdef extern const's
value from the C header at build time and ignores the literal written in the
.pxd, so one table is correct for every ta-lib release with no version
detection.

• 'NONE': -1 stays a Python literal — TA_FUNC_UNST_NONE no longer exists in
  current headers.
• 'ADXR', 'MFI' and 'STOCHRSI' remain accepted keys but are now no-ops
  that emit a DeprecationWarning; their enumerators are TA_FUNC_UNST_UNUSED_*
  in current ta-lib. They are deliberately not aliased to 'ADX'/'RSI' —
  that would turn a long-standing no-op into a call with side effects on every
  other function using ADX or RSI internally.
• _ta_lib.pxd: fixed the TA_FUNC_UNST_HD_PHASOR typo, dropped the stale
  literal values, and dropped the enumerators that no longer exist.

Test

test_unstable_period_moves_its_own_function covers all 20 real ids: setting an
unstable period of 5 must push exactly 5 more leading NaNs — 5 more leading
zeros for the one integer output — into that function's own output, and the
values both runs emit must be identical. Plus test_unstable_period_all, and a
warning/no-op test for the three deprecated names.

A get/set round-trip assertion cannot catch this class of bug; there's a comment
in the test saying so. Against unfixed master the new tests fail exactly where
the numbering diverges: ids 0–11 pass, 12 and up fail.

Verification

Built from source (setup.py build_ext --inplace, Cython 3.2.9, no wheels)
against three ta-lib C versions, same unchanged Python source:

ta-lib C	result
0.7.1	70 passed, 0 failed
0.6.4	69 passed, 1 failed — test_talib_version asserts b'0.7.1'
0.8.1 (dev build)	48 passed, 1 failed (same one), plus 2 uncollectable files (below)

TA_FUNC_UNST_ALL resolves to 24 on 0.6.x/0.7.x and 65535 on 0.8.x, and the
same source picks up both — that is what a corrected-constants patch could not do.

The 0.8.1 collection errors are pre-existing and unrelated; unmodified master
fails identically. _abstract.pxi's __get_flags raises KeyError: 33554432
on TA_FUNC_FLG_STREAM, a newer function-flag bit missing from its hardcoded
TA_FUNC_FLAGS dict — the same class of bug as this one, in another table.

talib/_ta_lib.c is left untouched, since it's only committed on Cython version
bumps.