fix(pir): a matching attribute name is not a successful read

[aboccag]

PaddlePirParser::GetOpAttr sets its found flag as soon as an attribute's
name matches, before looking at its type. When the type is one the overload
does not read, the caller's variable is left exactly as it was — uninitialised,
for the bare float alpha_; members mappers declare — and the enforce below
still passes, so nothing is reported.

pd_op.leaky_relu is the case that surfaced it. PIR stores its negative_slope
as a double; LeakyReluMapper asks for a float; the
isa<FloatAttribute>() test fails, nothing is written, and every exported
LeakyRelu gets

alpha = 2.466285297211678e-43     instead of 0.1

That is a denormal — whatever was on the stack — so leaky_relu becomes a plain
ReLU. The models convert, load and run; they are simply wrong. On
yolov3_mobilenet_v1_270e_coco the whole neck diverges (max|diff| 3.8 at the
first neck output, against 3.8e-06 through the entire backbone) and the detector
returns one box where Paddle returns 98.

What this changes

• Read a double where a float is asked for and vice versa. PIR stores as a
  double several attributes the legacy IR stored as a float.
• Accept integer array attributes in GetOpAttr<std::vector<float>>: Paddle 3.3
  stores some float-typed array attributes (full_int_array's values, for one)
  as integers.
• Make a type that cannot be read fail as loudly as a missing attribute,
  across the whole family of overloads rather than just the one.

That last point is a behaviour change, so it was measured rather than assumed:
across all 149 PaddleDetection detection configs, zero status changes —
every model that exported and converted before still does, model by model.

It also immediately caught a second instance: FullIntArrayMapper read dtype
into a std::string while the attribute is a DataTypeAttribute. It never used
the value — Opset7 takes the type from the output — so the read is dropped.

Verification

yolov3_mobilenet_v1_270e_coco after the fix, against Paddle inference over 50
COCO images: 136 boxes compared, worst pairwise IoU 0.999994, worst score delta
2.3e-06.

The fix also repaired three models that had been failing numerically for
unrelated-looking reasons — picodet_shufflenetv2_1x_416_coco,
centernet_shufflenetv2_140e_coco and ttfnet_darknet53_1x_coco, whose graphs
contain 36, 36 and 52 LeakyRelu nodes respectively. ttfnet_darknet53 had been
returning no boxes at all.

[CLAassistant]

All committers have signed the CLA.