[Express:Bugfix] Fix loadMap/load output readMap returning NULL

[fshuang8299]

Description

Variable::load / Variable::loadMap 加载的模型，输入可写（#4733 修复后）、推理可算，
但输出 readMap 返回 NULL——expressDemo 修复 #4733 后不崩但仍无输出。

Root cause：Variable::load 对非 INPUT 节点执行 Tensor::clone(allTensors[i])，
clone 共享源 describe（memoryType 默认 MEMORY_BACKEND），覆盖了 Expr::create
正确设置的 MEMORY_HOST。mapOutput 拷贝分支（allocMemoryForHostTensor 要求
MEMORY_HOST）无法分配 host → readMap 返回 NULL。

Fix：clone 后恢复 memoryType = MEMORY_HOST（与 #4733 的 INPUT 分支同源修复）。

// express/Expr.cpp（Variable::load 的 clone 分支内）
expr->inside()->mOutputTensors[index] = Tensor::clone(allTensors[outputIndex].get());
// 还原 Expr::create 的正确状态（memoryType=MEMORY_HOST）
TensorUtils::getDescribe(expr->inside()->mOutputTensors[index])->memoryType = Tensor::InsideDescribe::MEMORY_HOST;

Module

Express

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

Reproduction

模型由代码生成，无需附件（完整可编译示例见 issue #4750）。核心流程：

Variable::save({y}, "tiny.mnn");              // 生成 Conv3x3+ReLU 模型（输入 1x3x8x8 NCHW）
auto varMap = Variable::loadMap("tiny.mnn");
auto io = Variable::getInputAndOutput(varMap);
auto input = io.first.begin()->second, output = io.second.begin()->second;
auto ptr = input->writeMap<float>();          // ✅ OK（#4733 后）
// 写输入 0.5...
auto out = output->readMap<float>();          // ❌ 修复前 NULL → 修复后有效值

Verification

• ✅ 修复前后对比（临时还原 B1 实测）：readMap NULL → 1.36（数学正确：27×0.1×0.5+0.01）
• ✅ 新增回归测试 expr/LoadMapOutputReadMap 通过（独立验证）
• ✅ expr/ 全量单测 54 个通过（53 原有 + 新增 1）
• ✅ Module API / Session API 正常（1.36，无副作用）
• ✅ Const / 重复 readMap / resize 专项通过
• ✅ 量化模型 readMap 非 NULL（量化输出数值精度为量化链路本身问题，与本 PR 无关）
• 注：MNN 全量测试套件在 x86 批跑有卡点（op/InterpInt8、op/randomuniform 单独跑通过但批跑卡）——已用原始 master 对照实验确认与本次修改无关

Scope

本 PR 严格限定 Express Variable::load 的输出路径（1 行修复 + 回归测试）。
不影响：Module API / Session API / Const 节点 / 输入路径（#4733 已处理）。

Related

• Fixes Express loadMap/load 输出读取失败：output VARP 的 readMap 返回 NULL #4750
• 同源修复：[Express:Bugfix] Fix loadMap input tensor losing host buffer #4733（INPUT 分支，已合入）

[CLAassistant]

All committers have signed the CLA.

[wangzhaode]

Thanks for the fix. Could you please make two cleanup changes before merging?

1. Please rebase this branch onto the latest master and drop commit c47c693d from [Express:Bugfix] Fix loadMap input tensor losing host buffer #4733, since that change has already been merged. This PR should contain only the additional output readMap() fix for Express loadMap/load 输出读取失败：output VARP 的 readMap 返回 NULL #4750.
2. These two small loadMap regression cases do not need separate test source files. After rebasing, test/expr/LoadMapInputTest.cpp already exists on master; please add the output readMap() regression case to that existing file (or consolidate both cases into one common LoadMap test file) and remove test/expr/LoadMapOutputTest.cpp.

The functional fix itself looks reasonable. Thanks!

[fshuang8299]

Thanks for the review! Both cleanup changes are done:

1. Rebased onto the latest master and dropped c47c693 (from [Express:Bugfix] Fix loadMap input tensor losing host buffer #4733, already merged). This PR now contains only the output readMap() fix for Express loadMap/load 输出读取失败：output VARP 的 readMap 返回 NULL #4750.

2. Consolidated tests: added the expr/LoadMapOutputReadMap regression case into the existing test/expr/LoadMapInputTest.cpp and removed LoadMapOutputTest.cpp.

Verified after rebase: expr/ suite passes (54 cases, including both expr/LoadMapInputWriteMap and expr/LoadMapOutputReadMap).