【代码贡献】Fix SPSA optimizer ignoring parameter bounds

[mnn31]

Relates to #13.

Problem

spsa_minimize silently ignores the bounds argument in the normal case, and
corrupts it in the other case. The first path is what QAOA.run(optimizer='SPSA')
hits, since it always passes one pair per parameter; the second is reachable
through the documented spsa_minimize(..., bounds=[(min, max)]) public API.

Root cause

pyqpanda_alg/QAOA/spsa.py, _check_bounds (lines 17-33).

1. One (min, max) pair per variable, len(bounds) == d: the branch validates
   the dimension and then falls off the end of the function, so it returns
   None. _jail_inside reads None as "no bounds" and never clips.
2. One pair for all variables, len(bounds) == 1: return bounds * d
   multiplies the numpy array element-wise by d instead of repeating the row
   d times. _check_bounds(3, [(0, 1)]) gives [[0, 3]], which is the wrong
   interval and the wrong shape, so the later np.clip broadcasts one bogus
   pair over every variable.

Fix

Return the validated array in the per-variable branch, and use np.tile for
the single-pair branch. The empty-bounds warning and both existing
ValueError / IndexError checks are unchanged.

Verification

import numpy as np
from pyqpanda_alg.QAOA.spsa import spsa_minimize, _check_bounds

np.random.seed(0)
f = lambda x: float(np.sum(x ** 2) - np.sum(x))   # unconstrained min at x = 0.5
print(spsa_minimize(f, np.array([5.0, 5.0]), bounds=[(1, 2), (1, 2)], maxiter=200))
print(_check_bounds(3, [(0, 1)]))

before

[0.5 0.5]                  <- outside [1, 2]
[[0 3]]

after

[1. 1.]
[[0 1]
 [0 1]
 [0 1]]

Same thing through QAOA.run(optimizer='SPSA') with
gamma_bounds = beta_bounds = [(0.0, 0.5)]: on develop the returned
parameters come back outside the bounds; with this change they stay inside.
With the default (unbounded) settings the result is bit-identical to develop.

Tests

5 cases added to test/QAOA/Test_spsa_minimize.py: bounds respected end to end
for per-variable bounds and for a single tiled pair, _check_bounds returning
the array instead of None, single-pair tiling, and the existing validation
errors. Seeded, deterministic. 4 of the 5 fail on develop and pass with this
change.

pytest test/QAOA/Test_spsa_minimize.py     6 passed
cd test && pytest                          22 passed

The gradient line in spsa_minimize divides by xp - xm rather than
2 * ck * delta. With delta in {-1, 1} and no clipping these are
identical, and when clipping is active the current form is a finite difference
over the points actually evaluated, so I left it alone. Degenerate
lo == hi bounds make that denominator zero; that is a property of the
existing gradient line, not of this change, so it is out of scope here too.