#1076 fix(build): qualify size_t in vmm.hpp (GCC 12 rejects the unqualified name)

Beschreibung

## Problem

`include/strata/core/vmm.hpp:36` uses an **unqualified** `size_t`:

```cpp
bool mapped(int64_t i) const { return i >= 0 && i < chunks() && h_[(size_t) i] != 0; }
```

The header provides for itself with only `<cstdint>` and `<vector>`, and on GCC 12.2 (Debian 12's libstdc++) neither brings `::size_t` into scope. `src/core/vmm.cpp` includes this header first, so the **CUDA 12 engine build stops at 87%**:

```
[ 87%] Building CXX object CMakeFiles/strata_engine.dir/src/core/vmm.cpp.o
/opt/strata-0140/include/strata/core/vmm.hpp:36:73: error: 'size_t' was not declared in this scope;
                                                 did you mean 'std::size_t'?
```

`src/core/vmm.cpp` is **new in 0.1.40** (added at the release tip `1735d64`). 0.1.39 built fine on the same machine and toolchain.

## Environment (where this reproduced)

- Debian 12 container, `g++ (Debian 12.2.0-14+deb12u1) 12.2.0`, CUDA Toolkit 12.8, CMake 4.4.3
- The documented CUDA 12 build: `-DSTRATA_EXPERIMENTAL_SM60=ON -DCMAKE_CUDA_ARCHITECTURES="70;60" -DSTRATA_ENABLE_CUDA=ON`
- `set(CMAKE_CXX_STANDARD 20)` (top-level default)

## Minimal reproduction

Standalone, both C++17 and C++20 (so it is not a language-version issue):

```cpp
#include <cstdint>
#include <vector>
struct X { std::vector<unsigned long long> h_;
  bool mapped(long long i) const { return i >= 0 && h_[(size_t) i] != 0; } };
```

```
g++ 12.2: error: 'size_t' was not declared in this scope; did you mean 'std::size_t'?
```

## Fix

```diff
-    bool mapped(int64_t i) const { return i >= 0 && i < chunks() && h_[(size_t) i] != 0; }
+    bool mapped(int64_t i) const { return i >= 0 && i < chunks() && h_[(std::size_t) i] != 0; }
```

One token, compile-only: no behavior change. `expert_cache.hpp` already uses the `std::size_t` spelling, so this matches the house style; adding `<cstddef>` would also fix it, but the qualified name keeps the diff minimal.

## Verified

- V100 (sm_70) + P100 (sm_60) box, the CUDA 12 / SM60 build recipe above: **configure rc=0, build rc=0** with this change; binary carries 58× `sm_60` + 58× `sm_70` cubins; engine 0.1.40 has been serving since (`strata generate: session is up (engine 0.1.40)`).
- The unpatched tree fails at `vmm.cpp` (87%), the patched tree passes it — nothing else in the file or its callers changed.

Related but distinct: #1071 (missing `CUDART_VERSION` guard in the same new file for toolkits < 12.5) — this one hits newer toolkits too, so the two together cover CUDA 12.0–12.8 source builds.