Fix stack exhaustion DoS in json parsing via iterative shape traversal

[deezsecc]

Description

This PR fixes a stack exhaustion / Denial of Service (DoS) vulnerability in the TensorFlow Serving REST API JSON parsing logic.

---

Root Cause

When parsing incoming model prediction requests, rapidjson processes the request document iteratively. However, the post-parsing step in GetDenseTensorShape() (inside tensorflow_serving/util/json_tensor.cc) recursively traversed nested arrays (val[0]) to compute the tensor's shape without enforcing a recursion depth guard or limit.

An unauthenticated attacker could send a payload with a deeply nested array structure (e.g., 50,000 nested layers like [[[[... [1] ...]]]]), causing a stack overflow (SIGSEGV) and instantly crashing the tensorflow_model_server process.

A secondary recursive execution vector also existed in FillTensorProto(), which recursively parses nested elements up to the tensor's rank without validating that the rank is within safe limits.

---

Fix

This patch mitigates the vulnerability on both vectors:

Vector	Fix
Iterative Traversal	Refactored GetDenseTensorShape() to be fully flat and iterative, using a loop and pointer to resolve shape — eliminating recursion at the shape-calculation level
Rank Capping	Added a hard limit (kMaxTensorRank = 254, aligned with TensorFlow's TensorShape::MaxDimensions()) inside GetDenseTensorShape() to stop adding dimensions beyond safe nesting depth
Early Rejection	Added an early check in FillTensorProto() to reject payloads with rank > 254 via errors::InvalidArgument, preventing deep recursion on the value-filling path

---

Verification

• Iterative while loop behaves identically to the original recursion on standard payloads
• Deeply nested structures are capped early and rejected via FillTensorProto propagating InvalidArgument instead of exhausting the thread stack

[deezsecc]

@ch1nlu @rtg0795 Please review the PR. Thank you!