[QNN EP] Support htp_reused_io_limit_mb provider option for HTP AOT context load

[chunghow-qti]

Description

• Adds a new htp_reused_io_limit_mb provider option for the HTP backend, wired through to
  QNN_HTP_CONTEXT_CONFIG_OPTION_REUSED_IO_LIMIT when loading a cached QNN context binary (LoadCachedQnnContextFromBuffer).
• Usage: set ep.qnnexecutionprovider.reused_io_limit_mb value in MB. Default is 0, which leaves the SDK's default (unbounded) behavior unchanged.
• This option allow client specify the effective I/O buffer size used by a context instead of QNN HTP default estimation. QNN HTP uses this value for memory estimation and DSP PD placement decisions when it loads a context binary.

Motivation and Context

QNN HTP assumes no I/O reuse and estimates the context's memory using the total I/O size of all graphs in the context. If you actually share (reuse) one or more I/O buffers across multiple graphs in a context, or across multiple contexts, that default estimate can overshoot what the context really needs.

This overestimation matters because HTP loads each context into one of several process domains (PDs), each with a limited memory budget. An inflated estimate can cause a context to be placed in its own PD instead of sharing one with others (which is slower, since contexts on different PDs pay extra cost to talk to each other) — or it can cause the context to fail to load at all (QNN error 1002, "Failed to find available PD") even though it would actually fit.

By explicitly specifying the reused I/O size, you let QNN HTP use a smaller, more accurate estimate instead.

Test plan

• Context that doesn't fit in a PD → fits with the option
  Take a context binary that fails to load by default (QNN error 1002, "Failed to find available PD"). Load it
  • (a) without htp_reused_io_limit_mb → expect 1002;
  • (b) with a known-safe value → expect load success and correct inference output.
• Verify htp_reused_io_limit_mb is applied on both AOT context-load paths
  • contextCreateFromBinary: load four static-1024 contexts with
    htp_reused_io_limit_mb=2; confirm QNN verbose log reports
    Setting Reused IO memory estimation to 2 MB and all contexts load successfully.
  • contextCreateFromBinaryListAsync: run
    QnnHTPBackendTests.HtpSharedResourceOptimization_HtpReusedIoLimitMb_LoadsSucceeds;
    confirm QNN verbose log reports
    Setting Reused IO memory estimation to 128 MB.

[chunghow-qti]

/ci

[huaychou]

Do we need to add htp_reused_io_limit_mb in onnxruntime_perf_test?

[chunghow-qti]

will rebase on PR#778 and PR#677
Also updated the doc from owners info.

[chunghow-qti]

Update the doc to use htp_share_resource_optimization to clarify default path or not, which is more clear to user instead of using QNN API explaination.