intel_npu: reject oversized blob size before allocation in allocate_aligned_tensor (#37186)

## Summary

In `allocate_aligned_tensor` (Intel NPU plugin,
`src/plugins/intel_npu/src/plugin/src/blob_format_importers.cpp`), the
bounds check guarding against `blobSize` values too large to fit in a
`std::streamsize` was performed **after** the `ov::Tensor` allocation.
This is the bug reported in #37185.

On platforms where `std::streamsize` is narrower than `size_t` (e.g.
32-bit `streamsize`, 64-bit `size_t`), the allocation may succeed for an
oversized `blobSize`, after which the caller's `stream.read(...,
static_cast<std::streamsize>(blobSize))` wraps the value to a negative
`streamsize`, producing a truncated or no-op read instead of an error.

## Fix

Move the guard before the allocation so oversized sizes are rejected
immediately, before any resource is allocated:

```cpp
ov::Tensor allocate_aligned_tensor(size_t blobSize) {
    ov::Allocator customAllocator{utils::AlignedAllocator{utils::STANDARD_PAGE_SIZE}};
    if (blobSize > static_cast<decltype(blobSize)>(std::numeric_limits<std::streamsize>::max())) {
        OPENVINO_THROW("Blob size is too large to be represented on a std::streamsize!");
    }
    ov::Tensor tensor(ov::element::u8, ov::Shape{blobSize}, customAllocator);

    return tensor;
}
```

To make the function unit-testable, its declaration was moved from the
anonymous namespace into the `intel_npu` namespace in
`blob_format_importers.hpp`. Behavior is unchanged for callers.

## Test plan

Added regression tests to
`src/plugins/intel_npu/tests/unit/npu/blob_format_importers.cpp`:
- `AllocateAlignedTensorRejectsOversizedSize` — an oversized size (above
`std::numeric_limits<std::streamsize>::max()`) throws.
- `AllocateAlignedTensorAcceptsValidSize` — a valid size is allocated
without throwing.

## Verification

A standalone reproduction of the pre/post logic confirms the buggy
ordering attempted the allocation before the guard (and the
`static_cast` wrapped to a negative value), while the fixed ordering
rejects the oversized size before any allocation.

Fixes: #37185

---------

Signed-off-by: Deepak Bhagat <deepak988088@gmail.com>
Co-authored-by: Razvan Apetroaie <razvan-mihai.apetroaie@intel.com>

Author: deepak0x

src/plugins/intel_npu/src/plugin/src/blob_format_importers.cpp

@@ -20,6 +20,16 @@ namespace {
 
 using namespace intel_npu;
 
+ov::Tensor allocate_aligned_tensor(size_t blobSize) {
+    ov::Allocator customAllocator{utils::AlignedAllocator{utils::STANDARD_PAGE_SIZE}};
+    if (blobSize > static_cast<decltype(blobSize)>(std::numeric_limits<std::streamsize>::max())) {
+        OPENVINO_THROW("Blob size is too large to be represented on a std::streamsize!");
+    }
+    ov::Tensor tensor(ov::element::u8, ov::Shape{blobSize}, customAllocator);
+
+    return tensor;
+}
+
 constexpr std::string_view HANDLER_FACTOR_LOGGER_NAME = "blob_format_importer_factory";
 constexpr std::string_view RAW_BLOB_HANDLER_LOGGER_NAME = "RawBlobImporter";
 constexpr std::string_view BLOB_V1_HADNLER_LOGGER_NAME = "BlobFormatV1Importer";
@@ -35,16 +45,6 @@ constexpr std::string_view DECRYPTING_PAYLOAD_MESSAGE = "Decrypting the compiler
 
 const std::vector<size_t> CONSTANT_NODE_DUMMY_SHAPE{1};
 
-ov::Tensor allocate_aligned_tensor(size_t blobSize) {
-    ov::Allocator customAllocator{utils::AlignedAllocator{utils::STANDARD_PAGE_SIZE}};
-    ov::Tensor tensor(ov::element::u8, ov::Shape{blobSize}, customAllocator);
-    if (blobSize > static_cast<decltype(blobSize)>(std::numeric_limits<std::streamsize>::max())) {
-        OPENVINO_THROW("Blob size is too large to be represented on a std::streamsize!");
-    }
-
-    return tensor;
-}
-
 /**
  * @brief Special case for PERF_COUNT as it requires compiler_type detection in case it is still set to PREFER_PLUGIN
  */

src/plugins/intel_npu/tests/unit/npu/blob_format_importers.cpp

@@ -6,6 +6,7 @@
 
 #include <gtest/gtest.h>
 
+#include <limits>
 #include <sstream>
 #include <string_view>