[GPU] Fix ReduceMean axis element type (#34848)
## Details
Regression introduced by PR #34777 where affected models are as
following:
- common-sign-language-0002
- mobilenet-v2-1.0-224
- deeplabv3
- semantic-segmentation-adas-0001
- se-resnet-50
Those models fail on dGPU with:
`Axis 12884901890 out of the tensor rank range [-5, 4].`
## Root Cause
- `ConvertAvgPoolingToReduce` (dGPU-only, gated by `supports_immad`)
creates a `ReduceMean` axes constant as `element::i64`. This
transformation runs inside the main pass manager, but after the
`ConvertPrecision(i64→i32)` pass has already been registered and
executed within the same manager. The axes constant therefore survives
as i64.
- PR #34777 added a blanket `ConvertPrecision({i64→i32})` pass after
`UnrollTensorIterator`. This second pass converts the `ReduceMean` axes
constant's underlying storage from i64 to i32. However, `ReduceMean`'s
shape inference still reads the axes data as `i64`, reinterpreting two
adjacent `i32` values `[2, 3]` as one `i64` value `0x00000003'00000002 =
12884901890`, which exceeds the valid axis range.
- This only affects dGPU because `ConvertAvgPoolingToReduce` is
registered only when `device_info.supports_immad == true`.
## How to fix
- Change `ConvertAvgPoolingToReduce` to create the axes constant as
`element::i32` instead of `element::i64`.
- This aligns with the GPU pipeline convention that all integer
constants should be `i32`, and eliminates the `i64` constant that was
being corrupted by the post-unroll precision conversion pass.
## Tickets:
- CVS-183375
